Context Ontology for Software: Make Your Product the Tool Other AI Agents Call, With One Definition of Every Metric

Context Layer for AI Agents
Semantic Layer for AI
Tenant Isolation
Context Ontology Accelerator
AWS Context

Context ontology for software gives your product a governed layer of meaning that AI agents can rely on: one definition of ARR, seats, plans, entitlements, usage and tenants, plus the rules on who may see what. Your in-product agents read it, and so do customers' agents that call your product over MCP, so both return the answer your finance team would. Mactores builds it with your product, data and platform leads, runs it in your AWS account, first on the Context Ontology Accelerator and later on AWS Context once AWS opens that service, and puts the fee and ship date in the SOW before work starts.

Ask your in-product copilot what a customer's ARR was last quarter and you can get three answers. The CRM counts bookings, billing counts recognized run-rate and product analytics counts active accounts, so the agent reports whichever table it reaches first. Now picture a customer's own agent calling your product over MCP with the same question, or asking which of its workspaces went over their seat licence. Without a shared definition and a tenant boundary enforced by policy, that agent either guesses or, worse, joins across customers. Becoming the tool other agents call only pays off if your answers hold up when they call.

Start with one number your agents should never get wrong

Bring the metric, its competing definitions and an agent's answer. In a single session we locate the split, scope a first release and give you a realistic ship date.

Book a scoping call →

0product disruptions while Poshmark moved, stage by stage, onto a consolidated AWS platform
12 weeksfor Safaricom's multi-agent customer service to start delivering value
94%of Mactores engagements shipped by the committed date or sooner
90 daysof free fixes for production defects once you accept a release
2008when Mactores shipped its first work on AWS

AWS Agentic AI Specialization
AWS Premier Tier Services Partner
AWS Machine Learning Consulting Competency

Why Does Your In-Product Agent Report a Different ARR Than Finance?

Because each system that touches revenue defines it differently. The CRM counts bookings, billing counts recognized run-rate and product analytics counts active accounts, while discounts, credits and one-time services land in different places. An agent querying whichever table it finds inherits that disagreement. A context ontology holds one governed definition, so every agent returns finance's number.

The same split runs through every metric a software business lives by. A seat is a licence in billing, an invited user in the identity system and a weekly active user in analytics. A customer is an account in the CRM, a tenant in the platform and a billing entity with three subsidiaries. Churn depends on whether a downgrade counts. Text-to-SQL agents and retrieval pipelines resolve none of this on their own: they pick a table, guess a join and answer fluently.

Software teams are shipping agents faster than they are giving them context:

59.4%

of enterprise RAG architectures are still limited to basic or hybrid search, according to Futurum's August 2026 analysis, which also found 20% of respondents interested in agentic AI already run MCP in production.

Almost half

of companies in ICONIQ's July 2026 State of AI report say their agents still need a person on at least 30% of tasks, with lack of context or memory among the most common failure modes.

110M+

MCP SDK downloads a month by April 2026, as the protocol became shared infrastructure across 170 member organizations of the Agentic AI Foundation.

Those figures turn into the same handful of failures across product companies:

Where it breaks	Root cause	What the ontology fixes
The copilot reports ARR that finance won't sign	CRM, billing and analytics each define revenue their own way	One governed ARR, retention and churn definition, owned by finance
Seat and usage answers disagree with the invoice	Seats, users and metered events are counted in different systems under different rules	Plans, entitlements and metered dimensions modeled once and tied to billing
A customer's agent sees another customer's data	Tenant scoping depends on the agent writing the right filter	Tenant boundaries enforced by policy on every query, not by prompt
Support answers quote the wrong SLA	Plan terms sit in PDFs and contracts the agent can't connect to the account	Plans, contract terms and support tiers linked to each account
Nobody can say where an answer came from	No record of the definition, source and policy behind it	A per-answer trace your security reviewers can inspect

The pressure comes from customers as much as from boards. In August 2025, Gartner forecast that 40% of enterprise applications would include task-specific AI agents by the end of 2026, up from under 5%, and gave software leaders a three- to six-month window to settle their agentic product strategy. The products customers' agents prefer to call will be the ones whose answers match the customer's own books.

What Does a Context Ontology Give Your Product's Agents and Your Customers' Agents?

Your product's agents get one governed definition of every metric, entitlement and plan, plus tenant boundaries they cannot cross. Your customers' agents get the same context through a tool they call over MCP, so an answer is identical whether it comes from your copilot or theirs. Mactores models customers, tenants, plans, usage and rules, and writes access as policy.

We scope from the questions that matter commercially: the ones your copilot answers for customers, the ones customers' agents will send through your API, and the ones finance and RevOps put to both. Ten to 20 of them, each with an owner and an agreed answer, decide what goes into the first release, and the release is judged against them before anything ships.

Entity	What the ontology records	Where it lives today
Customer and account	The paying organization, its subsidiaries and billing entities, with one identity across systems	CRM account, billing customer, platform org ID
Tenant and workspace	The isolation boundary for data and access, and the workspaces or projects inside it	Platform database, identity provider, admin console
User and role	People, service accounts and agents, with roles and permissions per tenant	Identity provider, role tables, SCIM feeds
Plan and entitlement	What each tier includes: features, limits, seats and support terms, with effective dates	Pricing pages, billing catalog, feature flags, contracts
Subscription	Term, currency, discounts, credits and renewal dates	Billing system, quoting tool, contract repository
Usage event	Metered actions by dimension, with test traffic and retries flagged	Event pipeline, product analytics, metering service
Metric	ARR, net and gross retention, churn, activation and seat utilization, each with one formula and owner	dbt models, BI tools, finance spreadsheets
Rule and policy	Tenant isolation, data residency, plan limits and data-use terms	Policies, data processing agreements, code paths, contracts

Built and handed over in the first release:

A context ontology for your product, running in your AWS account and built for the question list your owners signed off
One customer identity across CRM, billing and platform, with the tenant and workspace hierarchy beneath it
Governed metrics for the ARR, retention, churn and usage figures in scope, reconciled to finance's report
Tenant isolation, plan limits and data-use rules expressed as policies and checks
An MCP tool your customers' agents can call, scoped per tenant and documented
A known-answer test suite that runs on every publish, with runbooks and stewards trained to use it

Drafted by agents

Proposed entities and relationships read from CRM, billing, platform and event schemas
Customer and tenant matches across systems, each with a confidence score
Draft metric formulas read back from dbt models and finance spreadsheets
Plan and entitlement rules extracted from the billing catalog and contracts
Regression runs against every metric on each change

Decided by your product, data and finance owners with our FDEs

Which ARR definition the board and the product both use
Whether a downgrade counts as churn or contraction
What a seat is when users, guests and service accounts all sign in
Which data a customer's agent may see, and through which tool
When answers are good enough to expose outside the company

That division mirrors how Mactores delivers every program and how the accelerator itself works; AWS describes its drafting step as turning months of ontology authoring into days. For a software company, the reclaimed time goes to the conversation that usually gets postponed: finance, RevOps and product agreeing, once, what ARR means.

How Does This Fit With the dbt Models and Metrics You Already Have?

Most software companies already run dbt models and a metrics layer for BI. Keep them. The ontology reads those definitions instead of duplicating them, then adds what a metrics layer never held: which customer owns which tenant, what each plan entitles, which data a given agent may touch and which rules apply. Apache Ossie, an open specification for semantic layers and ontologies that entered the Apache Incubator in July 2026 with more than 50 organizations behind it, is narrowing the gap between the two, and we map to it wherever your tools support it.

What Can a Context Layer for AI Agents Answer Inside and Outside Your Product?

With a context layer for AI agents in place, both sides get answers they can act on: which workspaces exceeded their seat licence, what NRR the enterprise segment posted on the board definition, which support SLA applies to a ticket, and why an invoice rose. Each answer stays inside the caller's tenant and carries its definition and source.

Customer-facing questions come first because they carry the most risk. A wrong answer inside your product reaches the customer directly, and a leak across tenants ends deals. Internal RevOps and finance questions follow on the same model.

Question	Who asks	What has to be true
Which workspaces in my account exceeded their licensed seats last month, and who administers them?	A customer admin, through your copilot or their own agent	What a seat is, and that the answer never leaves that tenant
What was NRR for the enterprise segment last quarter on the board definition?	Finance and RevOps	One NRR formula with cohort rules and exclusions, owned by finance
This ticket is from a Premium customer: which first-response SLA applies, and has it been breached?	Support staff and their AI assistant	Plan tier, contract overrides and timestamps linked to the ticket
Why did this customer's invoice rise 38% over last month, by metered dimension?	The customer's finance team or its agent	Usage events tied to price tiers, with test traffic and retries excluded
Which accounts on the legacy plan use features the new plan prices separately?	Product and pricing teams	Plan-to-feature entitlements mapped to actual usage
Which sellers listed heavily this week, share a payout account with a suspended seller and have open disputes?	Marketplace trust and safety	Relationships across sellers, accounts and disputes, with a person approving any suspension

Usage-based billing questions and marketplace trust and safety follow once customers, tenants and usage are settled, because they reuse the same identities and policies. The event pipelines, warehouse and streaming data these answers draw on come from our data platform modernization practice, which means the ontology starts from data that arrives on time and in a known shape.

How Does the Context Ontology Accelerator Turn Your Schemas Into Answers Agents Can Call?

It reads your schemas and documents, drafts an ontology your owners approve, and serves it through an MCP server and a SPARQL endpoint. Questions resolve to a governed metric where one exists, a graph query where joins are needed, and document search last. AWS intends these ontologies to carry into AWS Context, which has no release date yet.

The accelerator has been generally available since 31 July 2026 under the Apache 2.0 licence. It stores the graph in Amazon Neptune, indexes documents in Amazon OpenSearch Serverless and calls models through Amazon Bedrock, all inside your account, as AWS's announcement sets out. It is self-managed, carries no AWS Support and ships no software metric definitions, so ARR and churn are yours to define. If you plan to run it behind a commercial product, its open-source dependencies get a licence review first.

How Does One Tenant-Scoped Question Resolve?

Take a customer admin asking your copilot which of their workspaces went over the seat licence in September.

Tier one: a governed metric. Seat utilization per workspace is already defined, so the answer comes straight from it and no language model is involved.
Tier two: a graph query. To name each workspace's administrators, the agent queries a virtual knowledge graph in SPARQL, which reads your platform and identity tables without copying them.
Tier three: an agent over documents. If the customer's contract grants extra seats, the agent reads that clause under the same policies.

Throughout, the tenant comes from the authenticated caller, never from anything the agent writes. The accelerator's Cedar authorization and SQL firewall block cross-tenant joins and unbounded scans, and Amazon Bedrock AgentCore Policy, generally available since March 2026, applies the same Cedar language to the agent's tool calls.

What Do Scan, Model and Serve Involve for a Software Product?

Scan. Connect the sources your first questions need: the platform database, billing and CRM, the event pipeline and warehouse, dbt projects, and documents such as pricing pages, plan terms and support policies. Metadata is enriched automatically and candidate matches are flagged.
Model. Product, data and finance owners review the draft. Approved concepts become W3C OWL classes, with plan limits and isolation rules as SHACL shapes and R2RML mappings binding each class to its tables.
Serve. Your agents and customers' agents call the model through the MCP server, analysts and partners can use the SPARQL endpoint, and a tenant-scoped tool definition documents what each caller may ask.

Which Standards Does a Software Ontology Build On?

Software metrics have no formal standard, which is exactly why each company needs governed definitions of its own. These vocabularies supply the rest:

Standard	Covers	Role in your ontology
schema.org	Shared terms for products, offers, organizations, orders and software applications	Public-facing classes that customers' agents already recognize
Apache Ossie	An open specification for metrics, dimensions and relationships across semantic layers	Keeps governed metrics portable between your BI tools and the ontology
W3C ODRL	A policy language for permissions, prohibitions and duties	Plan entitlements and data-use terms written so agents can check them
Data Privacy Vocabulary (W3C community group)	Terms for personal data categories, purposes and legal bases	Data-use purposes attached to the data agents may touch
GS1 Web Vocabulary	Product attributes and identifiers such as GTINs	Catalog normalization for marketplaces and commerce platforms
Model Context Protocol	How agents discover and call tools	The route customers' agents use to reach your product's context

How Does Delivery Fit Into Your Release Train?

In five phases with signed exits: choose the questions, map customers, tenants and sources, agree definitions with product, data and finance, prove answers and isolation, then release behind a flag. Work runs alongside your release train, so ontology changes ship through the same review, testing and rollout you already use for code, never as a separate side project.

How we work sets the pattern for every Mactores program. For a software company, it means fitting into sprints rather than interrupting them. Definition sessions stay out of release weeks, and the ontology lives in a repository with pull requests, so your platform team reviews it the way it reviews everything else.

Phase 01
Choose the questions
What happens
Product, finance and support pick the 10 to 20 questions that matter most, name an owner for each, list where the data lives and record the expected answer; the ship date is set
Owner
Shared
Signed exit: SOW and question list countersigned
Phase 02
Map customers and tenants
What happens
Agents scan platform, billing, CRM and event schemas, and customer, tenant and user matches are reviewed with the people who own each system
Owner
Agents and FDEs
Signed exit: Identity map and tenant hierarchy approved
Phase 03
Settle the definitions
What happens
Finance, RevOps and product agree ARR, retention, churn and seat definitions; plan limits become checks; tenant and data-use policies are written with security
Owner
FDEs and your teams
Signed exit: Metric definitions and policies signed by their owners
Phase 04
Prove answers and isolation
What happens
Agents answer questions with known results, and cross-tenant probes and prompt-injection tests run against every exposed tool until each refusal is confirmed
Owner
Agents and FDEs
Signed exit: Accuracy and isolation results accepted
Phase 05
Release behind a flag
What happens
Agents switch to the ontology behind feature flags, customers' MCP access opens tenant by tenant, and runbooks pass to your on-call team
Owner
Shared
Signed exit: Customer-signed acceptance

Release gate for every agent, yours or a customer's:

Every metric in scope has one formula, one owner and a version
Every customer resolves to one identity across CRM, billing and platform
Known-answer questions meet the agreed pass rate
Governed ARR and retention match finance's last reported figures, or the differences are explained
Cross-tenant probes fail closed for every exposed tool
Prompt-injection and over-broad query tests pass
Every answer records its definition, sources, policy decision and ontology version
Open-source dependencies have cleared licence review if the ontology runs behind your product
The accelerator release is pinned and upgrades are rehearsed in staging
Cost alerts and per-tenant rate limits are in place

Scope, fee and ship date are tied together in the SOW at the end of phase one. If we cause a delay, we absorb it. If a delay starts on your side, such as late access to the billing system or an owner who can't make a phase review, extra hours are billed at the SOW's hourly rate. Read our commitment for the full terms.

What Has Mactores Shipped for Software Platforms That Can't Take Downtime?

Poshmark's platform moved to a consolidated AWS foundation in stages after its own team judged a clean cutover impossible, with every system and shared resource mapped first and no product disruption. Flipboard left a brittle Hadoop-backed foundation in weeks, with engagement rising afterward. Safaricom runs multi-agent customer service with firm limits on what agents may change.

Poshmark · Marketplace platform

A staged move its own team thought couldn't be done

Poshmark's systems depended on each other so tightly that its team concluded no cutover order could avoid breaking something downstream, and several integrators declined the work. Mactores mapped every system, integration and shared resource first, which revealed a staging order nobody had seen, then validated each stage against the live platform before switching it. The marketplace now runs on a consolidated AWS foundation, moved without disrupting the product.

Read the Poshmark story

Flipboard · Content platform

Off a brittle data layer without hurting engagement

Flipboard's Hadoop-backed database was fragile under engagement peaks, and the move away from it couldn't be allowed to hurt the product. Schema mapping and lineage across its data systems came first, and a parallel harness compared the new AWS-native foundation with the old one throughout. The work took weeks rather than quarters, and Flipboard's Greg Scallan credits the migration with better resilience and higher user engagement.

Read the Flipboard story

Safaricom · Customer service agents

Multi-agent service with hard limits on account changes

Scripted support flows at Safaricom broke whenever customers phrased things differently. Mactores built intent-based agents that work together across billing, accounts, products and troubleshooting, each allowed to change only what its guardrails permit. Value arrived within 12 weeks as routine requests left human queues. For a software product opening tools to outside agents, those authority limits are the template.

Read the Safaricom story

As Poshmark co-founder and CTO Gautam Golwala put it: "A big thank-you to the Mactores team for the execution." None of these was an ontology build, but each depended on what a context ontology depends on: an exact map of how systems relate and a clear line around what automation may do. Browse more on the Mactores stories page.

What Will Enterprise Security Reviews Ask About Your Agent Features?

They will ask how tenant data stays separated, whether customer data trains any model, what an agent can do without a person, how prompts and tool calls are logged, and how model or definition changes are approved. A context ontology answers most of these with evidence: tenant policies, per-answer traces, versioned definitions and tested refusals.

Enterprise buyers now send AI questionnaires alongside the usual security review, and the frameworks behind them are still settling. SOC 2 has no AI-specific trust services criteria yet, so, as a Cloud Security Alliance research note from August 2026 describes, auditors assemble their own requests, often asking for per-inference logs, model change records and evidence of how third-party model providers handle data. Two products can hold clean SOC 2 reports while having tested very different AI controls, which is why buyers increasingly ask to see the controls themselves.

Isolation by policy. The tenant comes from the authenticated principal, Cedar policies decide what each caller may reach, and the SQL firewall rejects cross-tenant joins and unbounded scans on every path, documents included.
Agents that ask before acting. Read-only answers can be automatic. Anything that changes a customer's account, billing or data waits for a person.
Definitions under change control. Metric and policy changes ship through pull requests, steward approval and the known-answer suite.
Evidence per answer. Each answer keeps the caller, tenant, definition version, sources, policy decision and model used, ready for a customer's auditor.

Requirement	What buyers and regulators expect	How the ontology helps
SOC 2 Type II	Evidence of access control, change management and monitoring for AI features, assembled by your auditor without AI-specific criteria to lean on	Per-answer logs, versioned definitions and policy test results
ISO/IEC 42001	An AI management system covering risk, oversight and lifecycle controls	Documented ownership, testing and change control for each definition and agent
EU AI Act, Article 50	From 2 August 2026, people must be told when they interact with an AI system, and generated content must be marked, with a grace period to 2 December 2026 for marking on systems already on the market	Disclosure and marking recorded alongside each agent interaction
California CPPA regulations	Risk assessments since 1 January 2026, and automated decision-making technology rules from 1 January 2027	The data each agent uses, and its purpose, recorded per answer
GDPR and data residency	Lawful basis, purpose limits and storage location for personal data	Data-use purposes attached to classes, with deployment in the Region each customer requires
Customer DPAs and contracts	No training on customer data without consent, and transparency about subprocessors	Data-use terms modeled as policies agents check before reading

What Can You Show a Customer's Security Reviewer?

Reviewer question	Evidence you can hand over
Could another customer's agent have read our data?	Cross-tenant probe results and the policy behind each exposed tool
Which definition of an active user produced the number in our quarterly review?	Definition history and approvals for that metric
What did your agent do in our workspace last Tuesday?	The per-answer log, filtered to that tenant
Is our data used to train anything?	Data-use policies and the model configuration on record
Who approved the change to the seat-limit rule?	The pull request, steward approval and test results

Which Software Businesses Need a Context Layer First?

Businesses whose customers already want to reach them through agents. B2B platforms with seat and usage pricing start with revenue and entitlement questions. Marketplaces start with trust and safety. Developer and data platforms start with usage and billing. Consumer platforms start with support and content. Each begins with the answer a wrong number would hurt most.

B2B software platforms

Seat, usage and hybrid pricing make ARR, retention and seat utilization the first governed metrics. The copilot inside the product answers from them, and customers' agents reach the same answers through an MCP tool, which is where AI for SaaS stops being a feature and becomes a channel. Products moving toward a headless SaaS model, where most use comes through APIs and agents, gain the most because every call depends on shared definitions. Tenant policies come first, tested and signed, before any external access opens to a single customer. Internet and software at Mactores

Marketplaces and commerce

Sellers, buyers, listings, payouts and disputes form the first model. Trust and safety agents follow relationships that keyword search can't, such as sellers sharing a payout account with a suspended one, and a person approves every suspension. Catalog normalization uses schema.org and GS1 vocabularies so listings carry the attributes buyers and their agents filter on. Seller-facing analytics then answer from the same definitions the finance team uses, so a seller and your own staff never see two versions of one number. See our internet and software work

Developer and data platforms

Usage-based billing leads. Metered dimensions, price tiers, projects and API keys link to each invoice line, so an agent can explain why a bill rose and which workloads drove it, with test traffic and retries excluded. Customers ask these questions through support today and will ask through their own agents next. Rate limits and per-tenant cost alerts keep agent access from becoming its own billing surprise, for you or for the customer. Software platforms at Mactores

Consumer and media platforms

Subscriptions, entitlements by region, content and support tiers lead. Support agents answer from the subscriber's real plan and history instead of a generic help article, and recommendation and engagement metrics get one definition across product and finance. Peak traffic shapes the design, so the most common questions resolve from governed metrics without a model call. Content rights and regional licensing questions join in a later release, on the same model and the same policies. Explore our internet and software practice

Fintech and health technology products usually combine one of these patterns with the stricter controls their regulators expect.

What Does a Context Layer Engagement Cost to Build, and to Run on AWS?

Most first releases ship in 8 to 12 weeks with 10 to 20 questions. Price reflects the questions, products and source systems involved, the gap between customer and tenant records, the metrics to reconcile with finance and whether customers' agents will call the ontology. Mactores sets fee and ship date when phase one closes.

Pricing note. Nothing here comes from a price list. Effort tracks your product's complexity: how many systems know about a customer, how many metrics finance reports and how many tenants and plans the policies must cover. The accelerator has no licence cost. The factors below explain a quote; your signed SOW fixes it.

Cost factor	Why it matters
Questions in scope	Each needs its own entities, links, rules and tests
Products and source systems	More products, billing systems and event pipelines mean more mappings
Identity quality	Customers, tenants and users that disagree across CRM, billing and platform need owner review
Metrics to reconcile	ARR, retention and usage definitions must match finance before agents use them
External exposure	Opening an MCP tool to customers' agents adds isolation testing, rate limits and documentation
Platform	Operating the self-managed accelerator means owning hardening, upgrades and on-call yourself or through us

Run cost. You pay AWS for graph storage, search capacity and model calls, and external access adds traffic you can now price. Metric questions resolve from governed definitions without a model call, which keeps per-request cost low enough to include agent access in a paid plan. Per-tenant rate limits and overnight shutdown of development stacks keep the bill predictable.

Which Context Layer Terms Should Product and Data Leaders Know?

Context ontology: The machine-readable description of your product and business that agents consult first: what exists, how things relate, the agreed meaning of each term and the rules that apply.
Context layer for AI agents: The shared set of definitions, relationships and policies every agent reads, so answers agree across agents, products and callers.
Semantic layer: A layer that defines business metrics and dimensions once for BI tools and queries, usually without tenant, entitlement or policy rules.
Metrics layer: The part of a semantic layer that holds metric formulas, such as dbt MetricFlow definitions.
Annual recurring revenue (ARR): The yearly value of recurring subscription revenue. Companies define it differently, which is why it needs one governed formula.
Net revenue retention (NRR): Revenue from a starting customer cohort a year later, including expansion and net of contraction and churn, as a share of where it began.
Entitlement: What a customer's plan allows, such as features, seat counts, usage limits and support terms.
Tenant isolation: Keeping each customer's data and actions separate on a shared platform, enforced by policy rather than by application code alone.
Headless SaaS: A product that exposes its functions to other software and agents through APIs and tools, with its own interface optional.
Text-to-SQL: Turning a natural-language question into a SQL query. Accuracy drops when joins, definitions and tenant boundaries are left for the model to guess.
Apache Ossie: An open specification, in the Apache Incubator since July 2026, for describing metrics, dimensions and relationships across semantic layers and ontologies.
Model Context Protocol (MCP): A protocol, now stewarded by the Agentic AI Foundation, that lets agents discover and call tools. Customers' agents reach your product's context through it.
Cedar: An open policy language, used by the accelerator and Amazon Bedrock AgentCore, for stating who may do what to which resource.
FDE: A Mactores specialist who joins your product, data and platform teams for the whole program, chairs the definition sessions and is on the hook for delivery.

How Can Mactores Help You Out?

Pick the entry point that fits where you are; each has a fixed fee and a contracted ship date. Copilot live but distrusted? We review its answers first. Questions clear, answers inconsistent? We build a first-release context ontology. Customers asking to call you from their agents? We design the MCP tool and its tenant policies alongside the ontology.

Since 2008, Mactores has shipped on AWS for product companies that cannot pause: marketplaces moved in stages, content platforms migrated under live load. That habit carries into ontology work. Definitions ship like code, isolation is tested like security, and no release lands during a code freeze.

Starting point	Engagement	Result
Your copilot is live, but customers or finance question its numbers	Agent answer review	A ranked list of the context each wrong answer lacked, plus a dated first-release scope
You know the questions, and answers still vary by agent	First-release context ontology	Agents inside and outside your product answering from one model your stewards maintain
Customers want their agents to call your product	Agent-ready product design	A tenant-scoped MCP tool backed by the ontology, with usage you can meter and price

Context ontology works the same way in every industry Mactores serves, and our context ontology services page covers the method and how we choose between platforms. Software companies tend to move fastest, because definitions already live in code and dbt, and they gain the most from going further. Once ARR, seats and entitlements are governed, the same model can price agent access, power customer-facing analytics and answer the RevOps questions that today take a spreadsheet and a meeting. It also gives your product team a clean contract to publish, so partners integrate against definitions rather than table names.

When the agents themselves still need building, or your product needs agent-native interfaces in place of forms and clicks, our AI agent development services handle that work on the same ontology and the same commercial terms. The ontology and everything around it stays yours, in W3C formats inside your AWS account, and we train your own people to extend it. Teams that start with one product usually add the next within a quarter, reusing the customer identity and tenant policies the first release settled, which is where the second release gets cheaper than the first.

Credentials matter once customers' security teams start asking, and all of ours are verifiable on our AWS partnership page: the AWS Machine Learning Consulting Competency, the AWS Agentic AI Specialization and AWS Premier Tier Services Partner status.

What Do CTOs Ask Before They Commit?

What is context ontology for software?

It is a model of your product and business, covering customers, tenants, users, plans, entitlements, usage, metrics and the rules that govern them, which AI agents consult before answering. It lets your own agents and your customers' agents give the same answer to the same question, inside the right tenant, with the definition behind it on record.

What is a context layer for AI agents?

It is the shared set of definitions, relationships and policies that sits between agents and your data, so every agent reads meaning instead of raw tables. A context ontology is one way to build it, and the most complete one, because it carries relationships and rules as well as metric definitions.

Is AWS Context available yet?

Not as of early November 2026. AWS announced it in June 2026 without a release date. Anything you build on the Context Ontology Accelerator is designed to move across, so there is little reason to wait: the slow work is agreeing definitions with finance and product, and that work transfers.

Do we need an ontology if we already have a semantic layer?

A semantic layer defines metrics and dimensions for queries and dashboards. An ontology also describes the entities behind them, how they relate, and the rules and permissions that apply, which agents need for questions that go beyond a single metric. Most software companies keep their semantic layer and let the ontology read its definitions.

Why do text-to-SQL agents get product metrics wrong?

Because they have to guess three things: which tables to join, which definition of a metric to use and which tenant's rows to include. A context ontology removes the guessing. Known metrics resolve from governed definitions, joins follow modeled relationships, and the tenant comes from the caller's identity rather than the generated query.

How do customers' agents call our product safely?

Through an MCP tool backed by the ontology. The caller's identity sets the tenant, Cedar policies decide which questions and data each caller may reach, rate limits cap usage, and every call is logged. Read-only questions can be answered automatically, while anything that changes the customer's account waits for a person.

Will the ontology replace our dbt models?

Yes. Keep the dbt models and metric definitions your analysts rely on. The ontology reads them rather than replacing them, then adds customers, tenants, entitlements and policies around them. Where your tools support Apache Ossie, definitions can move between the two without being rewritten.

Can we run the Context Ontology Accelerator behind our product?

Yes, with two cautions. It is self-managed open source with no AWS Support, so someone must own releases and on-call, and its dependencies need a licence review before anything reaches customers. Mactores handles both, or prepares your platform team to.

How is tenant isolation enforced?

By policy on every path. The tenant is bound from the authenticated principal, Cedar policies restrict each caller to its own data, the SQL firewall blocks cross-tenant joins and unbounded scans, and document search follows the same rules. Cross-tenant probes run before every release, and any failure blocks it.

Which metrics should we govern first?

Start with the ones agents already answer and finance already reports: ARR or revenue run-rate, net and gross retention, churn, seats or active users, and usage by billable dimension. Those are the numbers most likely to reach a customer, a board deck or an invoice, so they repay governance fastest.

When can agents start answering from it?

A first release usually ships 8 to 12 weeks after phase one begins, covering 10 to 20 questions on one product. More products, several billing systems or external MCP access extend that. The ship date goes into the SOW when phase one closes, and each later release is scheduled through the same step.

Does this work for consumer platforms and marketplaces?

Yes. Marketplaces model sellers, buyers, listings, payouts and disputes, and consumer platforms model subscriptions, regional entitlements and support tiers. The approach is the same, and the first release targets the questions that touch customers or revenue most often.

Who owns the ontology?

You do. The ontology and its mappings, policies, tests and runbooks sit in an AWS account and repositories you control; ownership, handover and exit terms go into the SOW before work begins. Your stewards learn to maintain it, so nothing depends on Mactores after handover.

Which Metric Do Your Product and Finance Teams Disagree On?

Send it our way. We'll show why the numbers split and commit to a date for settling it.

01
Take an FDE through the metric, the systems that calculate it and what your agents return, in one session.
02
Within a week, get our read, a first-release scope and the ship date we would sign up to.
03
Countersign the fixed-fee SOW and phase one begins.

A 30-day satisfaction guarantee applies to every engagement, and we fix, free of charge, any production defect raised in the 90 days following acceptance.

Settle my metric definitions →lets-talk@mactores.com