Intelligence Manager
One person accountable for a company’s AI, from what it costs to what it returns. Most companies now hire for this work in pieces, under several titles, often before they know they need one owner for all of it.
The problem.
A company now buys a third input beside people and capital: machine intelligence, sold by the token or by the seat. A token measures how much text a model read and wrote. A seat measures how many people have access. Neither one says whether the work was any good or whether it earned money.
Most of the value in a general-purpose model has to be found by the people using it: which work to give it, how to change the process around it, and how to tell a good answer from a bad one. Because the return is hard to see, many companies cap their AI budget to limit the downside, and leave value unclaimed.
The fix is one person whose job is to make the return visible and then raise it.
Titles it covers.
Companies post this work under several titles, often before they know they need one owner for all of it. A company that hires these separately has several people sharing one budget, with no single person answerable for what the spend returned. The Intelligence Manager is that person.
- AI Operations Lead
- Choose which workflows to automate, build the automation, and track workflows automated, test pass rates and cost per run.
- AI Platform and AI FinOps
- Model routing, caching, local models, token spend analysis, cost limits and budgets.
- Forward Deployed Engineer
- Work with the people who do the job, then build and ship agents connected to the company’s data and systems.
- Head of AI and AI Transformation
- Set AI priorities, choose vendors and pricing models, and report the return on AI spend to leadership.
- AI Adoption and Enablement
- Write the playbooks and train the people who use the tools, so a working use case spreads to other teams.
- AI Governance
- Guardrails, human approval, handoffs, audit trails, and license and data controls.
What the role does.
- Study the spend
- Pull usage and billing data from every AI vendor and internal model server, break it down by team, use case and model, and tie each use to the result it is meant to produce.
- Decide what to use it on
- Rank candidate workflows by value and by how clearly the result can be counted. Start with closed loops, where one interaction ends in a result that can be checked, such as a case resolved or a bill paid.
- Split the work
- For each step in a workflow, choose plain software, a model, or a person. Stop or shrink uses that cost more than they return.
- Make it efficient
- Send each request to the smallest model that passes its quality test, cache repeated prompt content, batch work that can wait, trim the context to what the task needs, and run open models in house where accuracy holds. Re-test before every model or prompt change.
- Build and deploy
- Map how the work is done today with the people who do it, then build the agent connected to the systems of record, ground prices and policies in tools, and build the handoff to a person. Release in stages, starting with a person approving each result.
- Measure the results
- Record each interaction’s result, keep a fixed test set per workflow, and publish cost per result, completion rate and handoff rate. Track adoption separately from impact.
- Advise on vendors and pricing
- Match each use to a pricing model and bring cost-per-result data to contract renewals.
- Spread the skill
- Find the staff who already use AI well, back them, and write a playbook for each working use case so other teams can repeat it.
- Report to leadership
- Send a regular report of cost per result by use case, and recommend where to raise or cut the budget, with the reason.
How it is measured.
- Cost per result
- Model spend, plus tool and hosting cost, plus human review time at a loaded hourly rate, divided by results completed.$ per result
- Tokens per result
- How much model work one outcome takes. It rises when prompts grow or retries climb.tokens per result
- Completion and handoff rate
- Share of interactions that end in the intended result with no person involved, and share passed to a person, broken down by reason.percent
- Test pass rate
- Share of a fixed test set the current release gets right.passed of total
- Cache hit rate and small-model share
- Share of input served from the prompt cache, and share of requests handled by the cheapest model that passes.percent
- Revenue attributed
- Revenue from projects that ran only because AI freed the staff time to run them.$ per quarter
Ways to pay for AI.
- Per token
- Fits when the company already knows how to apply AI and wants room to move work between uses. The buyer carries the risk.
- Per seat
- Fits broad staff access, where use is spread thin and hard to tie to one result.
- Per outcome
- Fits closed loops with a result that is easy to count, such as a resolved support case. The vendor carries most of the risk.
- Partnership
- Fits complex work with a large upside, where the vendor has expertise the company lacks. The risk is shared.
Why a company funds it.
- A known return
- Leadership sees cost per result for each use, so the budget grows where it pays and shrinks where it does not.
- Lower cost for the same work
- Routing, caching, batching and smaller context lower the bill, and every change is tested first so quality holds.
- Growth without matching headcount
- Shopify reported 34% revenue growth in Q2 2026 (SEC filing) while its COO describes hiring as flat. Block reported 25% gross profit growth in the same quarter (shareholder letter).
- New products
- Repeated cases that an agent and an expert work through together can be packaged and sold as a service.
- Better contracts
- Usage and result data give the company a basis to negotiate, and to choose outcome pricing or a partnership where those fit.
The first 90 days.
- Days 1 to 30
- Collect every AI bill and usage log, build the usage record by team and use case, and interview the heaviest users and the teams with the most repetitive work.
- Days 31 to 60
- Apply the first efficiency changes behind quality tests, choose one closed-loop workflow and build its agent with a person approving each result, and start recording results per interaction.
- Days 61 to 90
- Publish the first cost-per-result report, reduce the approval step where the tests show it is safe, and propose the next two workflows.
Where the idea comes from.
The role is drawn from Nate B Jones, “Enterprise AI Pricing: Why Everything You Know Is Wrong” (October 2026), and from the job postings companies use for each part of it. My measured results in this role are on the portfolio and in the experience page.