Contextual Agentic

Lab

What Does an AI Agent Cost to Own? A Five-Driver TCO Model

Cost the five drivers, turn released hours into a defensible benefit, compare buy, extend and build, and see ROI, payback and sensitivity.

Runs in the browserIntermediate

30 minutes with your own numbers; five minutes to see the shape·Chapter 17 — Evaluation, reliability, observability and replay; Chapter 20 — The enterprise model portfolio

What actually operates

These controls run for real. They are the point.

  • Cost by month, split into once and recurring — Each line item has a one-off cost paid before it runs and a monthly cost paid after. Payback and discounting need that split, and a yearly total hides it.
  • Costs that grow with use — Compute, storage, egress, telemetry and consumption grow with volume each year; people and licences stay flat until somebody decides otherwise.
  • Discounting — Net present value discounts each month's net cash flow at the rate you set. At a rate of zero it is plain net value.
  • Payback from cumulative cash — Payback is the first month in which cumulative benefit overtakes cumulative cost, with the up-front spend landing in month zero.
  • The hour is spent once — Released hours are allocated across avoided cost, extra throughput and service level. The shares cannot add to more than 100%, so the same hour cannot be counted twice.
  • Sensitivity, one variable at a time — Each driver, then the benefit, is raised and lowered by 20% on its own. A model that moves several things at once cannot be argued with.

What is simulated

Stand-ins. Nothing here touches a real system.

  • Every amount — There are no prices in this lab. Fields start empty. The example loads round, made-up figures so the shape is visible before you have your own, and is labelled illustrative wherever it appears.
  • The three paths in the example — Buy, extend and build are filled with invented numbers. Which one comes out ahead in the example is an accident of those numbers and says nothing about your organisation.
  • The gate reading — A simple test of net value, present value, your hurdles and a 20% benefit shortfall. A real committee also weighs strategy, risk appetite and things no spreadsheet captures.
Boundaries
  • This is a teaching model of a decision, not a prediction of your costs. It is educational material only.
  • It is not financial, procurement, legal, tax or professional advice, and it is not a substitute for a business case your finance function has reviewed.
  • No figure here is a vendor price, a benchmark result or a service-level commitment. Nothing on this page ranks or endorses any supplier or any of the three paths.
  • The gate reading applies only the hurdles you enter. It is a prompt for an investment committee, not a decision, and the illustrative example says nothing about which path suits your organisation.

Build your total cost of ownership

Fill in what you know, driver by driver. Leave out what you do not: the page shows what is missing instead of guessing. Add the value of the work and it works out return, payback and what would change the answer.

Nothing you type leaves this page. The model runs entirely in your browser: no amount and no result is sent anywhere, stored on a server or logged. The only copy kept is in your own browser until you clear it.

The five drivers and what they contain

Infrastructure

What it costs to run the thing: compute, somewhere to keep the data, moving it around, and keeping it up when something fails.

  • Inference compute (CPU and GPU). Hosted-model charges belong under consumption below. Put capacity you buy or reserve here.
  • Storage. Documents, indexes, vector stores, logs, traces and every copy kept for evidence.
  • Network and egress. Data leaving a cloud, crossing regions or reaching a private link. Easy to forget, hard to reverse.
  • Resiliency and high availability. Second region, failover, backups and the capacity that sits idle so a failure is survivable.
Development and integration

Building the agents and connecting them to the systems that hold the work. Mostly paid once, and almost always larger than the plan.

  • Agent design and build. Prompts, tools, tool contracts and the loop around the model.
  • Orchestration. Workflow, state, retries, human hand-offs and the runtime that holds them.
  • Connectors to existing systems. The integrations to systems of record. Usually the slowest and least predictable line.
  • Identity and authorisation. Agent identity, delegated authority, secrets handling and the policy points that enforce them.
  • Compliance and security testing. Assessments, penetration tests, audit evidence and the review time they consume.
Data quality and preparation

Making what the agent reads fit to be read, and keeping it that way. Data work recurs; treating it as a one-off is the commonest way to under-budget.

  • Cleaning and labelling. De-duplication, fixing, tagging and the review time of people who know the content.
  • Grounding and indexing. Chunking, embeddings, graph or index builds, and rebuilding them when the design changes.
  • Refresh schedules. Keeping the sources current. A stale index is a quiet failure that nothing alerts on.
  • Drift monitoring. Detecting that the data, the questions or the model's behaviour have moved.
Expertise and staffing

The people who design, run and change it, and the people who must work differently because of it. Usually the largest driver, and the one the platform price never shows.

  • Architects. Solution, data and security architecture time, including review boards.
  • AI and ML engineers. The people who build and tune the agent and its evaluation.
  • MLOps and platform engineering. Deployment, pipelines, environments and keeping releases boring.
  • Subject-matter experts. The time of the people who know what correct looks like, for labelling, review and acceptance.
  • Change management and training. Adoption is people changing how they work. Budget it, or the benefit never arrives.
Operations and licensing

Keeping it measured, correct and affordable once it is live, plus the charges that grow with use.

  • Monitoring and telemetry. Traces, metrics and logs, and the storage and tooling to read them.
  • Evaluation. Running the evaluation sets, refreshing them and the people who judge the edge cases.
  • Retuning and prompt maintenance. Models change, data changes, requirements change. This is maintenance, not a project.
  • Incident response. On-call, investigation, and the work of explaining to a regulator what happened.
  • Consumption and licences. Token charges, platform fees and per-seat licences. Often the line the plan was built around.

An agent business case usually shows the platform price and a saving. What decides whether it pays is everything in between: the people, the data work that never ends, the connectors, the controls and the running costs. This model makes you cost all five drivers, say where the released hours go, and compare the three ways of getting there before anyone commits.

How to use this, in order 1. **Set the horizon and the rates.** Three years is a common starting point. Leave growth and discount at zero to see plain totals first. 2. **Say what the work is worth.** Enter volume, adoption, minutes saved and a loaded labour rate, then allocate the released hours to where they go. If finance already has an agreed figure, use that instead. 3. **Cost one path at a time.** Open each of the five drivers and enter what you know, once and per month. Anything left empty counts as zero and is shown as a gap. 4. **Repeat for the other paths.** Buy, extend and build share the business case; they differ in cost shape, time to value and risk. 5. **Read the matrix, then the return.** Look at which driver is largest and how much is paid up front. 6. **Open the sensitivity view.** The longest bar is where your estimate needs the most evidence. 7. **Read the gate reading last.** It tells you which hurdle a path misses, not what to decide.
Time saved is not money saved An agent that saves 10,000 hours has created capacity, not a cheque. It becomes value only when the capacity goes somewhere, and there are three places it can go: - **Avoided cost.** A hire not made, overtime not paid, a contract not renewed. This is the one that reaches the income statement, and only if someone actually makes the decision. - **Extra throughput.** The same people handle more. This is value only if there is demand for more and each extra transaction earns something, so the model asks for contribution, not revenue. - **Service level.** Faster, better, fewer penalties. Real, but valued per hour by you, and the easiest to inflate. Each released hour goes to one of them or to none. The shares in this model cannot add to more than 100%, and what is left is shown as absorbed: it made people's weeks easier and is worth nothing to this case. A route only counts when you say you have evidence for it. The rest is reported as upside and left out, because a business case that includes everything it hopes for will be taken apart at the first review.
The five drivers, and what each one hides - **Infrastructure.** Inference compute, storage, network egress and the idle capacity that makes failure survivable. Hosted-model charges belong under consumption, not here. - **Development and integration.** Agent design, orchestration, connectors, identity and authorisation, and compliance testing. Connectors are usually the slowest and least predictable line. - **Data quality and preparation.** Cleaning, labelling, grounding, indexing, refresh schedules and drift monitoring. This is recurring work, and treating it as one-off is the commonest way to under-budget. - **Expertise and staffing.** Architects, AI and ML engineers, MLOps, subject-matter experts, change management and training. Often the largest driver, and the one no platform price shows. - **Operations and licensing.** Monitoring and telemetry, evaluation, retuning, incident response, and consumption. Models change, data changes and requirements change, so this is maintenance, not a project.
Buy, extend or build The three paths capture the same value at different costs and risks. - **Buy** an existing SaaS agent: light on development and staffing, heavy on licences, fastest to value, least control. The exit cost is easy to leave out. - **Extend** a copilot or platform you already run: reuses identity, data and controls you already pay for. Watch the licence uplift and the point where its limits push the work back onto custom code. - **Build** a custom agent: heaviest on development, data and people, most control, longest road to value. The recurring cost of owning it is the part business cases leave out. Compare them on the same horizon, the same growth and the same discount rate. Differences in time to value belong in "benefit arriving in year one", and differences in delivery risk belong in the risk haircut. Put the haircut on the path that carries the risk, not on all three.
Reading the answer **The matrix** shows each driver by year and its share of the total. A path whose cost sits mostly in the first column is a project; one that sits in the later columns is an operation, and needs an owner and a budget line that survive the project team being disbanded. **Payback** needs the split between up-front and recurring cost and the timing of the benefit. It cannot be found from an annual benefit and a first-year total alone: two plans with the same yearly figures can pay back months apart. **Net present value** matters when the horizon is longer than a year or two, because a benefit that arrives in year three is worth less than one that arrives in year one. **Sensitivity** tells you where to spend your evidence. If the benefit is the longest bar, prove the benefit before you refine a cost line. If one driver dominates, ask for a second quote.
What this model leaves out, and a committee will ask about - **Risk-adjusted value in more detail.** Failed runs, hallucinations, human review, model drift and remediation. The risk haircut is a single number; yours may need to be several. - **Model routing economics.** Frontier, smaller and local models and their fallbacks, compared per successfully completed task rather than per token. The platform economics lab does this. - **Control costs.** Security, audit, privacy, data residency, operational resilience and the cost of leaving a vendor. Enter them under the drivers that carry them, and enter exit cost as a one-off. - **Strategic and option value.** A capability you could not otherwise build, or a position you could not otherwise hold. This model values none of it, deliberately.
Four ways this goes wrong - **Counting freed hours as cash.** They are capacity until spend elsewhere falls. - **Counting the same hour twice.** As avoided cost and again as extra throughput. - **Costing the build and not the owning.** Data refresh, evaluation, retuning and incident response recur every month. - **Comparing a bought price with a built cost.** Cost the licence uplift, the integration and the exit on the bought path as carefully as the engineering on the built one.