Human AI Judgment
Work with me

Job vs AI · Sales

Account Manager

The only profession in this series where a completely correct output can still be the thing that costs you the account.

  • relationships
  • renewal
  • disclosure
  • commitments
Get the playbook

Everywhere else, the test of an AI output is whether it is right. Here a message can be entirely right and still do damage, because the person reading it can tell it was generated - and that changes where the line falls.

The problem

Start with the problem, not the tool

In most professions the test of an AI output is whether it is correct. Here a message can be correct in every detail and still cost you the account, because correctness was never the thing at risk.

Most AI advice for commercial work starts from a capability - personalise at scale, more touches per week - and looks for somewhere to apply it. Applied to relationship work that inverts the job: the customer is not buying volume of contact, they are buying somebody who pays attention.

The order that works runs problem, then workflow, then decision, then technology, then verification. This profession adds one gate the others do not have. Before anything is sent, ask whether you would be comfortable if the recipient knew exactly how it was produced. If the answer is no, that is the finding, not a detail to manage.

That one question separates most of the good ideas here from most of the bad ones, and it splits almost exactly along direction of travel: work that points inward at your own records is fair game, and work that points outward at the customer is not, unless a person owns the words.

Lane 01

Automate

Stable rules, checkable outputs, affordable failures - and all of it pointed inward at your own records.

  • Activity logging and record hygiene

    Writing calls, emails and next steps into the CRM without you retyping them. This is the one automation in the profession that improves the relationship, because a record somebody else can pick up is what makes cover possible.

  • Capturing and structuring meeting notes

    Transcription plus extraction of commitments, dates and owners. Conditional on telling people they are being recorded, which is a legal question in some places and a trust question everywhere.

  • Pipeline and renewal-date reporting

    Renewal dates, entitlements and contract terms already live in systems that can be queried exactly. Asking a model what a contract says is slower and less correct than reading the field.

  • Usage and entitlement reporting

    Assembling what the customer actually used against what they are paying for. Mechanical, checkable, and the backbone of a renewal conversation that is about outcomes rather than persuasion.

Lane 02

Augment

A draft or a wider search saves real time - and anything that will reach the customer is rewritten by a person first.

  • Preparing the brief before a meeting

    CRM history, product usage, ticket history, contract terms and prior notes assembled with every line traceable to its source. The strongest case in the profession, because nothing it produces leaves your desk.

  • Drafting a follow-up you then rewrite

    A first pass listing what was said and what you committed to, so you do not drop one. The line that references something personal is exactly the line that reveals the message was assembled rather than written.

  • Assembling the renewal case

    Usage, outcomes and open commitments turned into something the customer can recognise as their own year. Preparation, not persuasion.

  • Summarising a long support history

    Two years of tickets read down to what actually went wrong and how it was resolved, before a review where somebody will remember it better than you do.

Lane 03

Human

The relationship itself, and everything that commits the company to something. This is the largest lane here, and it is what the customer is paying for.

  • Negotiating price and terms

    A model can lay out the options and the history. It cannot know which concession buys goodwill and which one teaches them to ask again next year, and the person who says the number is the one who committed to it.

  • Deciding what to promise

    Every external commitment has to be one the organisation can honour. That is a judgement about capacity, priority and risk, and it belongs to somebody who is allowed to make it.

  • Noticing unhappiness before anyone writes it down

    Dissatisfaction reaches a system last. By the time it is a ticket, a low score or a non-renewal, the decision has usually already been made in a conversation you were not part of.

  • Handling a service failure

    An apology is only worth anything because a person chose to make it. Anything generated here is a well-written way of not making the call.

Lane 04

Verify

Unusually large here, and mostly not about hallucination. Nobody imposes these checks on you, which is why they are the ones that get skipped.

  • Every external claim, against what can actually be delivered

    Capability, price and timeline confirmed by somebody who can commit to them, on this customer's contract and this customer's version. The closest thing this profession has to an imposed check, and the imposition is contractual rather than professional.

  • Consent and preference state, before any outreach

    UK guidance is unambiguous that people have an absolute right to object to or opt out of direct marketing at any time. Where the rules apply, contacting somebody who objected is a compliance failure as well as a relationship one, and the message being accurate is not a defence.

  • Any predicted number, before it drives a decision

    At least one major vendor documents its own predictive scoring as blackbox machine learning: inputs and outputs known, the transformation between them not. Name the observable thing behind a score - usage down, tickets up, the champion left - or you have a number rather than a reason.

Situations

Seven situations, worked

Seven representative situations from account work, chosen because they land in different lanes. The counts below are counts of these examples and not a measurement of the profession.

  • 01Automate1example
  • 02Augment2examples
  • 03Human2examples
  • 04Verify2examples

  • Writing the follow-up after a call

    Augment
    dailyA draft that lists what you committed to is genuinely useful. Every sentence that reaches the customer is then rewritten by you, in your own words.

  • Preparing the brief for tomorrow's account review

    Augment
    per reviewThe safest strong case in the profession: it assembles several systems and produces nothing anybody outside the company ever sees.

  • Logging the call and keeping the record true

    Automate
    dailyCapture, structure and write back, with the commitments spot-checked. Telling people they are being recorded is not a formality.

  • Acting on a score that says the account is at risk

    Verify
    weeklyA prediction is a prompt to look, never a finding. If you cannot name the observable thing behind it, you have a number rather than a reason.

  • Answering whether the product can do what they just asked

    Verify
    weeklyThe risk is not that a model invents a capability. It is that you repeat it, in a meeting, to somebody who will hold you to it.

  • Negotiating the renewal price and terms

    Human
    annualAssembling the facts you negotiate against is useful preparation. The concession itself is a commitment somebody has to be allowed to make.

  • Noticing the account has gone quiet

    Human
    continuousNothing in the systems will tell you. A scheduled check-in message is worse than silence, because it shows the calendar prompted it rather than you.

Solutions

Choose the kind of solution, then the tool

Organised by the problem being solved. Named products are examples of a class, never a recommendation.

  • No technology at all

    The account is unhappy and needs a conversation, or a person to own it.

    Watch for: Building a report about somebody who wanted a phone call is the most common expensive mistake here.

  • Conventional software and documented APIs

    Getting exact account state out: renewal dates, entitlements, approved price lists, usage.

    For example: CRM object APIs, product usage data

    Watch for: Paying for intelligence where a query would have been exact, cheaper and reproducible.

  • Traditional machine learning

    Predicting propensity or churn from enough closed history of the same shape.

    Watch for: This is prediction, not generation, and at least one vendor documents its own model as a blackbox. A score nobody can explain should never be the reason a customer is given.

  • Generative drafting and extraction

    Briefs, summaries, commitment extraction, and first-pass follow-ups.

    Watch for: The moment the output is sent rather than rewritten, the class was chosen correctly and the boundary was not.

  • Agents for repeated multi-step assembly

    Pre-meeting briefs and commitment registers built from several read-only systems.

    Watch for: Worth it only when the setup cost amortises - and in this profession with one extra rule: whatever it does, it does not send.

Chatbot → agent

From chatbot to agent

A progression, not a leap - and here it is allowed to stop early more often than in most professions.

  1. Ask a chatbot to tighten a paragraph you wrote

    You bring the context, you take the output away, you rewrite it in your own voice. This is genuinely useful and it is where most account managers should start.

    Augment

  2. Save the prompt that works

    With your own voice and your approved claims in it. Many account managers should stop here: a chatbot with a good saved prompt is enough for almost all of the drafting this job needs.

    Augment

  3. Give it the real material

    Paste in the call notes and ask what you committed to. Still one turn, still entirely yours, and now grounded in what was actually said rather than what you remember.

    Augment

  4. Connect it to the CRM and the product data

    So it reads the current account rather than what you pasted. This is the step that turns a saved prompt into something that needs permissions, and read-only is the boundary.

    Automate

  5. Let it run the sequence, and check itself

    Pull, cross-check, list, flag - against an output contract you wrote. One extra rule applies here that applies nowhere else in the series: whatever it does, it does not send.

    Verify

Blueprint

What an account agent looks like

One worked blueprint. The playbook builds a second one that assembles every commitment made to an account, and explains which candidate was rejected and why.

  • Goal

    Produce a pre-meeting brief for one account: the situation, the stakeholders, the open commitments and the likely objections, each traceable to a record.

  • Rules

    Never state a capability or price that is not on the approved list. Distinguish public fact from inference. Never draft anything that will be sent, and never contact anybody.

  • Tools and memory

    CRM, product and support systems, read-only. Memory holds account history and who the stakeholders are - never a stored judgement about a person.

  • Output contract

    Situation, stakeholders, open commitments, usage trend and likely objections, each with a source link that opens to the ticket, clause or logged call it came from.

  • Verification

    Every line opens to its source record before you rely on it in the room. A brief you cannot trace is a set of assertions you are about to repeat out loud.

  • Escalation

    Anything implying a commercial commitment, any claim it cannot source, and any contradiction between what the contract says and what the usage data shows.

  • What it must not do

    Contact anyone, draft outbound copy, quote a price, or write to a source system. Everything it produces stops at your desk, which is what makes it defensible.

Risks

Where this goes wrong

Two kinds of failure. Verification catches only the first, and in this profession the second is the entire story.

  • Detectable automation

    High risk

    A message that is right in every particular and reads as generated. The recipient's conclusion is not that you used a tool; it is that the attention they thought they were getting was never there. No accuracy review catches it, because there is nothing wrong with the message.

    What to do:A person writes every word that reaches a customer. Use the test directly: would you be comfortable if they knew exactly how this was produced?

  • Disclosure and consent failures

    High risk

    Contacting somebody who objected, running an AI system that talks to a customer without telling them, or putting a note-taker in a call nobody was warned about. Where the relevant rules apply these are compliance failures, and the content being accurate is not a defence.

    What to do:Check preference state before any outreach, disclose where disclosure is required, and treat recording consent as a question with a real answer rather than a formality.

  • A capability or price claim the company cannot honour

    High risk

    The classic correctness failure, and it lands in a customer's inbox rather than in your drafts. Language models are fluent about product capability and have no access to your roadmap or this customer's contract.

    What to do:Every external claim is confirmed by somebody who can commit to it before it is made. "I will find out" costs thirty seconds; the alternative costs the renewal.

  • A score used as a reason

    Medium risk

    A propensity or churn number that decides who gets attention, quoted internally as though it were a finding and occasionally to the customer as though it were evidence. Vendor documentation may itself describe the model as a blackbox.

    What to do:Require the observable fact behind any score before it changes what anybody does, and never give a customer a model output as the reason for anything.

  • Customer data in a tool you are not permitted to use

    Medium risk

    Their contract with you, your data-protection obligations and your own policy all constrain where account data may go, independently of how well the tool performs.

    What to do:Settle the permission question before the trial. Discovering the answer afterwards is not a recoverable mistake with somebody else's data.

Sources

Research basis

What the profession framing and the risk positions rest on. The lane assignments above are Human AI Judgment analysis, not a classification made by any of these organisations.

+Playbook

The full analysis, in a form you can take into work.

The Practical AI Judgment Playbook: Account Management

Twenty-seven pages on where technology genuinely helps an account manager: seven situations worked through four lanes, two agent blueprints, and the places where a completely correct message still costs you the account.

Cover of The Practical AI Judgment Playbook: Account Management

In your own words. This decides which profession I analyse next.

Optional.

Optional. You will receive the requested playbook whether or not you opt in — this box only controls future emails.

Keep going

Other professions

The same structure, applied to different work.

  • Delivery

    Project Manager

    The profession whose actual deliverable is information other people act on - which makes it the one where faster reporting can quietly make the job worse.

  • Healthcare

    Nurse

    The profession where the honest answer is most often that technology should stay out of the way - and where the work it can take off you is worth more than anywhere else.