# What AI can now do for your most important customers
Date: 2026-11-10
Author: Richard Brooks
Canonical: https://richard-brooks.com/ai-for-key-accounts/
In March 2025 the best AI model could work on its own for about an hour. By January 2026 the figure was over five hours.
That change matters more to key account management than any chatbot feature. A strategic customer is a long job: years of history, a dozen people to keep track of and a plan that needs constant upkeep. Most of that work happens between the meetings, and most of it gets done at 1000 MPH.
A year ago AI helped an account manager write. It can now do a good share of the work in between. This is a guide to seven jobs it can take on, as questions and answers, with seven figures to try. Some of the figures are redrawn from the slides I teach from on [Cranfield's Key Account Management Best Practice programme](https://www.cranfield.ac.uk/som/kambp).
## What changed?
Two things. The first is stamina. METR, a research group, measures the length of task that a model can finish on its own about half the time, counted in the hours a person would need. In [March 2025](https://metr.org/blog/2025-03-19-measuring-ai-ability-to-complete-long-tasks/) the best model managed about an hour. Its [January 2026 update](https://metr.org/blog/2026-01-29-time-horizon-1-1/) put the best model it had measured at 320 minutes, with wide error bars, and found the figure doubling roughly every three months since 2024.
Figure 1
How long an AI can work on its own
METR's 50% time horizon: tasks of this length, measured in human working time, are completed about half the time. The March 2025 figure is for Claude 3.7 Sonnet. The January 2026 figure is for Claude Opus 4.5, with a confidence interval of 170 to 729 minutes.
The second is access. AI used to see only what somebody pasted into a chat box. It now connects to email, calendars, shared drives and the CRM. In August 2026, Salesforce and Anthropic [announced](https://investor.salesforce.com/news/news-details/2026/Salesforce-and-Anthropic-Announce-Claudeforce-The-1-AI-Meets-the-1-AI-CRM/default.aspx) 37 prebuilt sales skills that work on live customer data from inside Claude. An admin connects it once and every seller on the team has it.
So an agent can work for hours at a stretch, on the real record of the account. On the course I describe an agent as a very intelligent, well-motivated intern who doesn't know a lot about your business (yet). It gets a brief, access to what it needs, a budget and a deadline. Somebody checks its work.
## Where are we on the map?
Mark Davies and I drew five ages of AI in 2024, and I have taught from the slide ever since. We move the "we are here" marker every time we present it.
Figure 2
The five ages of AI
Redrawn from my Cranfield slide (November 2024, updated October 2026). Each age can be selected. The key account lines were written for this figure.
In the age of the prompt, a person asked and the machine answered. Everything reset when the chat closed. In the age of the agent, a person delegates a goal and the work runs on a schedule. The next age, the collaborator, is a piece of work that the person and the AI both keep improving. Most of the seven jobs below sit on the line between agent and collaborator.
## Can the account plan keep itself up to date?
The annual account plan has a familiar life. It gets written, presented and filed. Delegates on our programme tell me theirs lives in the filing cabinet and never gets looked at again.
Nobody has ever had the hours to keep a document current against a live account.
An agent has the hours. In April 2026 Andrej Karpathy published a short note, [llm-wiki](https://gist.github.com/karpathy/442a6bf555914893e9891c11519de94f), that fits the problem. It has three layers: raw sources, a wiki of pages the AI writes and maintains, and a schema that tells the AI how the wiki is structured. He calls the result "a persistent, compounding artifact".
For a key account the three layers are easy to name. The sources are emails, call transcripts, review notes and contracts. The pages are one per stakeholder, one per opportunity and a ledger of value delivered. The schema is the team's own account planning method, written down as rules.
Figure 3
A week in a living account plan
1 · Sources
2 · Living pages
What changed
3 · Schema. The team's account planning method, written as rules the AI follows on every pass.
An invented account. Three things arrive in a week. The AI files each one, updates the pages it touches and reports what changed, including one contradiction with the plan.
I keep my own contacts and markets this way. I don't maintain it. Scheduled agents check it daily and tidy it weekly, and every email, call or note I add makes the rest more useful. For an account manager, that could be one key account: every stakeholder, meeting, decision and contract term, all linked.
One rule keeps the memory trustworthy as it grows. One AI owns the writing. Everything else reads: other tools, other agents and the team.
## Does the relationship map match the evidence?
Most relationship maps record what the account manager believes. A list of contacts with a column headed "strength of relationship", filled in once.
An AI can read a year of correspondence in one pass and compare it with that column.
Figure 4
The relationship map and the record
An invented account with six contacts. The first view is the plan's own rating of each relationship. The second is what a year of emails and meeting notes shows.
The same pass shows who has gone quiet, who has started copying in procurement and which promises were never followed up. No account manager enjoys reading that. I would still rather read it in a plan than in a lost renewal.
## Can the team rehearse the room?
A buying group is several people who want different things. A model can play them.
On the course we run this as a 15-minute exercise. An invented biscuit maker pitches to three invented buyers at a supermarket: a category buyer, a procurement lead and a finance business partner. Each has a short CV that tells the AI who they are. Each raises one objection in their own voice, and the presenter answers them one at a time. Then the three talk it over and give a verdict: shortlist, more evidence, or reject. Last, each names the claim they trust least.
Figure 5
Pitch it to the buying group
Redrawn from a course exercise. Brooks Biscuits is invented and so are the three buyers. No real person's data is involved. The objections and questions were written for this figure.
The value is in what happens next. The AI steps out of character and turns each objection into a question for the real customer: who to ask, by role, how to ask it, and what answer would change the proposition. Every answer turns a guess into a confirmed figure.
I'd check the internal questions first. Asking the buyer something your own finance team already knows costs credibility.
## Can the value be put in the customer's own words?
A proposition written in the customer's vocabulary lands better, and their results call already says what that vocabulary is. The method has three steps, each a prompt.
1. **Extract.** "From this transcript, list every phrase where the CEO or CFO describes what they are trying to achieve or are worried about. Quote word for word, name the speaker, give the page."
2. **Compare.** "Here is our value proposition. Which of its words does the customer never use? Which of their priorities does it miss?"
3. **Rewrite.** "Rewrite it in the customer's own language. Change no claim and no figure. Mark every phrase you took from the transcript."
Figure 6
One proposition, before and after
Illustrative, from a course slide. Brooks Biscuits is invented. The claims and figures are the same in both versions. In the second, the customer's goal opens the sentence and the supplier's capability serves it.
The same reading produces the numbers. An agent will go through the customer's annual report and results in minutes and return their priorities, risks and areas of investment with sources. What used to take a day now takes minutes, and domain knowledge is what makes the output useful.
I'd keep two rules for any figure that reaches a customer. The AI proposes and ordinary code does the arithmetic. Every number says where it came from.
## Who watches the account between meetings?
Signals come in two kinds. Buying intent: hiring for roles the supplier's product supports, a mention of a vendor review, complaints about a current supplier, a budget announcement. Strategic intent: a new market, an acquisition, a change of leadership, a partnership, a sustainability commitment.
Nobody has time to look every day. An agent on a schedule does. Agents earn their keep on work that is repetitive, scheduled and mostly reading. It checks what it finds against the plan and leaves a short note when something matters.
They are the wrong tool for judgement calls, relationships and anything irreversible. In the systems I build, agents draft, file and remember. A person presses send.
## What happens when the account manager leaves?
In most firms the real knowledge of a key account sits in one person's head and inbox. When that person moves on, the customer has to explain themselves all over again.
A living plan built from the record holds that memory. A successor can ask it why the pricing changed in 2024, or who objected at the last renewal, and get an answer with the source attached.
It matters where the memory is kept. On 12 June 2026 a US government directive led Anthropic to [suspend two of its models](https://www.anthropic.com/news/fable-mythos-access), and my project chats on them were out of reach until access came back on 1 July. I keep the record in plain text files that I own. The work should outlive the app.
## Can something be built for one customer?
A pricing calculator. A shared dashboard for a joint project. A briefing page with the customer's strategic themes, who to contact, what to say and why now. Each of these used to need a developer and a budget.
Each is now an afternoon's work. Co-created value has always been the top rung of account management, and it used to be expensive.
## Isn't the customer doing this too?
Yes. Gartner [surveyed](https://www.gartner.com/en/newsroom/press-releases/2026-03-09-gartner-sales-survey-finds-67-percent-of-b2b-buyers-prefer-a-rep-free-experience) 646 B2B buyers in August and September 2025. 45% had used AI during a recent purchase.
Procurement got there early. Harvard Business Review [reported](https://hbr.org/2022/11/how-walmart-automated-supplier-negotiations) in 2022 that Walmart's negotiating chatbot was closing agreements with 68% of the suppliers it approached. At its 2025 symposium Gartner forecast that AI agents would handle [90% of B2B purchases](https://www.digitalcommerce360.com/2025/11/28/gartner-ai-agents-15-trillion-in-b2b-purchases-by-2028/) by 2028, worth more than $15 trillion.
So a proposal is now read by the buyer's AI as well as the buyer. A claim with no number and no source gives a machine nothing to read.
Figure 7
What a procurement bot can check
Three invented claims from an invented supplier, each in a vague and a quantified version. Selecting a claim switches it. An illustration of how a claim-scoring tool might read a proposal. The figures are made up.
What would a procurement bot make of your value proposition today?
## How would a team set this up?
Four building blocks, each answering a different question.
| Building block | The question it answers | For a key account |
| --- | --- | --- |
| Preferences | Who am I, and how do I like answers? | Set once. British English, short answers, my role. |
| A project | What does the AI need to know about this account? | One per account: the annual report, the results call, the account plan and its own instructions. |
| A skill | How is this job done, every time? | A written procedure the AI loads when the task matches: a review pack, a buying group map, a value case. |
| A connector | What is happening right now? | Live data from email, calendar, drive and CRM, where the company allows it. |
Three things need guarding. The data: public documents can go anywhere, and customer material goes only into a tool the company has approved. The instructions: an instruction runs in every chat, so a wrong one is wrong every time. The output: nothing reaches the customer that a person hasn't read.
## Where is the advantage?
Everyone can rent the same model, and the CRM vendors are handing out the skills. I see the advantage in two things a competitor can't copy.
The first is the account's own history. Years of emails, calls, contracts and delivery data. A firm that has kept it in good order gets far more from the same AI.
The second is the method. A team that has written down how it scores an account, maps a buying group and counts value can hand all of that to the agent as rules. I tell every class the same thing about strategy: [don't be very good at implementing a bad one](/dont-be-very-good-at-a-bad-strategy/). Agents multiply a good strategy and a bad one alike, so domain knowledge has never mattered more.
## Does this need a key account team?
No, and I think that is the biggest shift. Most of my readers run language companies. Plenty of them have five or ten customers who pay for everything else and nobody whose job is to manage them.
Proper account management for those customers used to need people a firm that size couldn't afford. An owner with a well-kept inbox and one agent can now keep a living plan on each of them.
## What stays human?
Trust, judgement and the decision about which customers deserve the investment in the first place.
The AI does the reading, the modelling and the watching. The account manager gets time back for the part the customer values most, which is being understood by a person.
If METR's trend holds, the working time will have doubled again by spring. An account's record takes longer than that to put in order, so the record is where I'd begin.
**Sources:** METR, "[Measuring AI ability to complete long tasks](https://metr.org/blog/2025-03-19-measuring-ai-ability-to-complete-long-tasks/)" (March 2025) and "[Time Horizon 1.1](https://metr.org/blog/2026-01-29-time-horizon-1-1/)" (January 2026); Salesforce and Anthropic, "[Claudeforce](https://investor.salesforce.com/news/news-details/2026/Salesforce-and-Anthropic-Announce-Claudeforce-The-1-AI-Meets-the-1-AI-CRM/default.aspx)" announcement (26 August 2026); Karpathy, A. "[llm-wiki](https://gist.github.com/karpathy/442a6bf555914893e9891c11519de94f)" (April 2026); Anthropic, "[Fable and Mythos access](https://www.anthropic.com/news/fable-mythos-access)" (2026); Gartner, "[Sales survey finds 67% of B2B buyers prefer a rep-free experience](https://www.gartner.com/en/newsroom/press-releases/2026-03-09-gartner-sales-survey-finds-67-percent-of-b2b-buyers-prefer-a-rep-free-experience)" (9 March 2026); Van Hoek, R. and others, "[How Walmart automated supplier negotiations](https://hbr.org/2022/11/how-walmart-automated-supplier-negotiations)", *Harvard Business Review* (November 2022); Digital Commerce 360, "[Gartner: AI agents will command $15 trillion in B2B purchases by 2028](https://www.digitalcommerce360.com/2025/11/28/gartner-ai-agents-15-trillion-in-b2b-purchases-by-2028/)" (28 November 2025); Cranfield School of Management, [Key Account Management Best Practice](https://www.cranfield.ac.uk/som/kambp).