An email at $10,000
A consultant at one of the top strategy firms told me last month how his employer manages his AI spend. He has unlimited access. When he passes $10,000 in a month, he gets an email. The firm's message was explicit: the threshold triggers a notification and nothing else.
Then he added: "I'm sure I cost more than I make."
The policy tells us what the firm permits. It doesn't tell us what he spends. But the arithmetic is easy to recognise. I have seen research skills at private equity funds that take three hours to run and cost a few hundred dollars in model usage each time. At $250 a run, twice a day, over 21 working days, one analyst reaches $10,500 a month before a follow-up question. The opportunity is to reuse context across the team through a company brain.
That is an illustration. I don't have an industry average to offer. It explains why I pay attention when research teams say they are always out of tokens. In the workflows I see, a substantial part of the bill buys context somebody else in the firm has already assembled.
The spending gap is real
Public data on AI spend per employee is patchy. Ramp's August 2026 AI Index offers one useful view: in July, the median business in its data spent $11.95 per employee on AI. The top 10% spent $650, and the top 1% had a median of $7,400. The last figure is more than 600 times the first.
The data covers US businesses on Ramp. It doesn't single out research firms, and it covers more than tokens. It shows how far heavy users have moved from the typical business in that dataset.
Reported policies show the response. PYMNTS reported in November that Uber had exhausted its annual AI budget by April. It also described a Tesla approval threshold of $200 a week, with an exception for xAI beta products, and an October reduction from $300 to $200 a day for some Apple engineers.
At Accenture, Forbes relayed reporting by 404 Media about $29 per employee per month in central Claude funding, with additional usage charged to projects. Accenture disputed the description of a firm monthly access limit. Each report describes a different policy at a different date, so the figures don't compare.
A spending threshold can make a bill visible. It cannot tell you whether the research was worth buying, or whether the firm bought it five times.
Why research firms feel it first
A consultant, analyst or portfolio manager spends much of the day building an understanding of a company, market or person. Every new assignment begins with gathering context.
The material is long: information memoranda, data rooms, expert-call transcripts and years of board packs. Reading it consumes tokens. Reading it again in the next session consumes them again.
The data can cost money twice. Firms pay for the model and for data vendors, some of which meter their AI connectors. I have watched a researcher use half of a provider's annual connector allowance in a weekend. A cheaper model would have reduced only one part of that bill.
Where the tokens go
Watch a research run step by step and you see how much of it is preparation.
The tools arrive before the question
When a client loads every connector's tool descriptions up front, those descriptions consume context on each request. I once added a connector and found that even a trivial question exceeded the context window. Modern tool discovery can avoid loading everything. The point is to inspect what your agent sends.
Documents pass through the model more often than they need to
Anthropic's code-execution example illustrates the cost. Passing a two-hour meeting transcript between tools through the model twice could add 50,000 tokens. In a separate worked comparison in the same post, loading tools on demand and handling data in code reduced usage from 150,000 tokens to 2,000. That example shows better tool design. It doesn't measure what a company brain saves.
Each session rebuilds the same understanding
On a complex company, an agent can spend half an hour assembling context. Then the person at the next desk asks about the same company and starts again. An investor at a venture fund told me she is always out of tokens and has no idea what her colleagues are researching. A fresh data room lands, and several people analyse the same material in parallel.
Failures and oversized contexts add to the bill
Retries can make a cheap-looking configuration expensive. On one of our pipelines, increasing the reasoning budget per step from 2,000 to 8,000 tokens reduced total spending because fewer runs failed. The right budget depends on the task.
More context can also make retrieval less reliable. Anthropic's context-engineering guide describes declining recall as context grows. Paying to include everything is no assurance that the model will use the right part.
Six ways a company brain reduces repeated work
A company brain is a knowledge graph built from the systems where the firm works: documents, email, meetings, CRM and licensed data. People and agents can ask it directly. The saving comes from doing the preparation once and letting the team reuse it.
1. Start from the research a colleague already did
Everything the firm knows about a company, person or project is gathered in one place, and findings can be saved back with their evidence, so the next analyst starts from existing work. New questions still need new information and further reasoning. They do not need to rediscover every relationship from raw files.
2. Add sources without bloating every request
Agents reach every connected source through one connection, so adding another source need not add another large catalogue of tools to every agent request. This reduces tool overhead; it does not make all token usage independent of how much data the firm holds.
3. Read only what the question needs
A search returns up to 100 short excerpts, narrowed by source and date, and the agent reads in full only the few it needs. Irrelevant material stays out of the model's context.
4. Reuse paid data where the licence allows
A vendor record can sit beside the firm's other evidence about a company, so later questions can use it without fetching it again. Reuse must follow the vendor's licence, the reader's access rights and the required refresh schedule. It saves repeat fetches. Keeping information current still costs money.
5. Pay less for routine processing
Classification, extraction and matching seldom need the most expensive model for every step, so the brain does not use one for them. On one lead-scoring engagement, we estimated that comparable analysis through a general-purpose model API would cost £40 to £50 per company, about ten times our cost. That estimate comes from one engagement, and we haven't published the comparison.
6. Get caching discounts on repeated prompts
Repeated prompts with a stable opening can qualify for provider caching discounts. As checked on 14 November 2026, Anthropic lists cache-hit input for Opus 5.5 and Sonnet 5.5 at 5% of their standard input rate. Prices and discounts vary by model, and cache writes have a separate cost.
A shared graph doesn't create a cache hit by itself. The provider's rules still apply: an identical eligible prefix, the relevant cache lifetime and the same permitted cache scope. Different organisations do not share Anthropic caches. Measure actual hits rather than assuming every colleague's request reuses one.
What changes, and what still costs money
A brain moves spending as well as reducing it. In most installs, historical ingestion is the cost peak: the archive has to be processed before people can reuse it. There are ongoing costs for updates, storage, retrieval, model calls and paid data.
The saving depends on overlap. If analysts never revisit the same companies or sources, shared preparation buys less. Deal, sector and portfolio teams often return to the same names. That is where I would look first.
We do not yet have a published benchmark. My estimate from client work is that centralising research context can remove double-digit percentages from a research team's total token spend. It is an unpublished estimate, and we don't promise it. We are building a public comparison.
Sharing also has a boundary. Each passage keeps its source permissions. An analyst cannot reuse a board pack they cannot read, and a company-level agent cannot search a colleague's private brain. Our permissions guide explains how those boundaries work.
A flat spending cap removes valuable runs alongside wasteful ones. First establish where repeated preparation sits in the bill. Then decide how much the remaining research is worth.
Run a two-week audit on one research team
Before buying anything, choose a team and a repeatable deliverable: an investment memo, a company screen or a client deck.
- Measure cost per finished deliverable. Include failed runs, model usage and paid tools. Record quality and turnaround time alongside the bill.
- Inspect a sample of runs. Separate context assembly, reasoning and output where the logs allow. Identify tool overhead, retries and provider cache hits.
- Count the overlap. How often did two people research the same company or document in the same week? Could existing work have answered the next question?
- Check paid-data usage. Compare connector calls with quotas, licences and refresh requirements. Add ingestion, updates and storage to any proposed savings calculation.
If repeated context dominates, test a shared brain against that baseline. If overlap is low, model routing or better tool discovery may be the better first change. Compare total cost at a quality level you accept. Tokens saved tell you little on their own.
Buy the research. Keep the understanding.
The aim is to make useful research available to the next person who needs it, with its evidence, dates and permissions intact. That also addresses the alignment problem: the firm can see and reuse work that would otherwise disappear into individual threads.
We install a first company brain use case in two weeks. Choose 60x Cloud, hosted by us with EU or US data residency, or Your Cloud, inside your own Azure, AWS or GCP cloud environment.
Book a demo and bring a research workflow your team runs more than once. We can work out what is being repeated and what a measured pilot should test.
Sources
- Ramp AI Index, August 2026, published 12 August. Spending data for US businesses on Ramp; not specific to research firms.
- PYMNTS, Tech Giants Ration Premium AI as Company Budgets Hold, 6 November 2026. Reported company policies, including reporting attributed to The Information.
- Forbes, Accenture Clarifies Claude Spending Limits as AI Cost Questions Grow, 16 October 2026. Includes Accenture's response to 404 Media's reporting.
- Anthropic, Code execution with MCP, 4 November 2025. Worked examples of tool and data-handling overhead.
- Anthropic, Effective context engineering for AI agents, 29 September 2025.
- Anthropic pricing and prompt-caching documentation, checked 14 November 2026. Pricing and cache rules can change.
Client and prospect examples are anonymised. Illustrations and engagement estimates are labelled throughout. 60x installs company brains for investment firms and professional services teams.