AI for Sales Operations: the spine behind the number
Let us start by ruling something out. The most common thing people mean when they say "AI for sales" is a machine writing cold emails at volume. That is a bad use of this technology. It floods inboxes, burns your sending domains, trains buyers to ignore you, and produces a pipeline of people who never wanted to talk. It also solves a problem you probably do not have. The bottleneck in most sales organisations is not the volume of messages leaving the building. It is that the machinery behind the sellers is held together by spreadsheets, chase emails and the memory of two long-serving people.
That machinery is where agents belong. Sales operations is high volume, deadline driven, rule governed and spread across systems that never quite agree: the CRM, the CPQ tool, the ERP, the billing platform, the commission spreadsheet, the contract repository, and a shared inbox where deal desk requests arrive with no structure at all. Nobody sees the whole picture on one screen. That is precisely the shape of work an agent handles well.
CRM hygiene and enrichment, done continuously
Every forecast, territory decision and commission calculation rests on CRM data that decays the moment it is entered. Accounts get duplicated because someone typed the legal entity name instead of the trading name. Contacts leave and nobody marks them inactive. Close dates sit weeks in the past. Industry codes and headcounts were filled in once, by whoever created the record.
An agent works this continuously rather than in the quarterly clean-up that never happens. It reads new and modified records, matches them against your master data and external registries you already licence, and proposes merges for duplicate accounts with the evidence attached: matching registration number, matching domain, overlapping contacts. High confidence merges you have authorised it to make, it makes, logging what it merged and why. Low confidence ones go to a person with both records side by side, which turns a twenty minute investigation into a ten second decision.
It fills gaps the same way. Missing parent company, missing sector classification, missing headcount band, missing billing entity. It writes back only fields you have designated as agent maintained, so a human's entry is never silently overwritten. And it flags contradictions rather than resolving them: when the CRM says one billing entity and the ERP says another, that is an escalation, not a guess.
Pipeline inspection and forecast roll-up
Pipeline inspection is the weekly ritual where a manager reads twenty opportunities and asks the same five questions about each. Is anything actually happening on this deal? Who is the economic buyer? Has the close date moved again? Does the stage match the evidence?
An agent can do the reading. For every open opportunity it pulls the activity history, the calendar record, the emails logged, the notes, the contract status and the stage history, and produces an inspection summary per deal: last meaningful customer contact, whether a named decision maker is engaged, how many times the close date has slipped and by how much, whether required stage criteria are actually satisfied, and which fields contradict the narrative. A deal at stage four with no meeting in five weeks and no procurement contact gets flagged. A close date that has moved three times gets flagged with the pattern shown.
Then it rolls the numbers up. The agent assembles the forecast by segment and territory, reconciles it against the previous submission, and writes a variance note: what came in, what slipped out, what was added, what changed value, and which movements were made after the submission deadline. That variance note is usually the single most useful artefact in the whole cycle, and it is the one nobody has time to write.
The important boundary: the agent does not decide the number. It prepares the evidence and the arithmetic so the humans who own the commitment argue about the right deals rather than about whose spreadsheet is correct.
Quote, pricing approval and the deal desk
Quoting is where sales operations spends its patience. A rep asks for a discount. Somebody has to work out whether the discount is within policy, what the resulting margin is, whether the customer already has a most favoured pricing clause, whether the term and payment schedule are standard, and who has to approve it.
An agent takes the quote request and does the checks before a human sees it. It reads the configuration, prices it against the current rate card and the customer's existing agreement, calculates the effective discount and margin, checks it against the approval matrix, and identifies every non-standard element: uplift caps below policy, payment terms beyond your standard, unusual termination rights, a service level the delivery team has not committed to before. It then routes to the correct approver with a one page brief rather than a bare request, and it chases approvals that go quiet.
For deal desk review of non-standard terms, the agent compares the proposed language to your playbook clause by clause, marks each as standard, pre-approved fallback or genuine exception, and pulls the precedent: when this clause was last accepted, on which deal, at what value, with which conditions. Legal and finance still decide. They simply stop starting from scratch.
Territory, quota and the order handoff
Territory and quota administration is a burst of intense, error-prone work at the start of every period, followed by a year of disputed edge cases. An agent can apply the rules: assign accounts by the segmentation criteria you defined, detect accounts that satisfy two territory rules or none, model the coverage and load balance, and produce the exception list for the humans who own the plan. Mid-year, when an account is reassigned, it traces the downstream effects on open pipeline, quota credit and commission and writes the change record so the dispute six months later has a documented answer.
At the other end, the order handoff. When a deal closes, an agent validates that the signed order is internally consistent before it reaches fulfilment and billing: signature present, entity matching the CRM account, quantities matching the quote, start date achievable, billing schedule mapped to the right products, purchase order attached where the customer requires one. Bad orders caught here cost minutes. Caught at first invoice they cost weeks and a credit note.
Commissions, renewals and churn signals
Commission calculation is arithmetic on contested inputs. Most disputes are not about the plan, they are about which quarter a deal landed in, whether a split was recorded, or whether a clawback applies to a cancelled subscription. An agent recalculates statements from source records rather than from the spreadsheet everyone has been editing, produces a per rep explanation showing the line items behind each figure, and pre-flags the cases likeliest to be challenged: deals closed within days of a period boundary, splits without written agreement, adjustments applied manually last cycle.
When a dispute is raised, it assembles the file: the opportunity history, the plan clause in question, the precedent from previous cycles, and its own reading of the outcome. A human decides. The value is that the decision takes ten minutes instead of a fortnight of email.
Renewal and churn signal detection is the same discipline pointed forward. The agent watches the things that actually precede a loss: support ticket volume and sentiment, usage decline, the champion's departure from the account, an unanswered renewal notice, invoices going past due, meetings declined. It scores accounts, writes the signal detail into the CRM where the account owner will see it, and opens a task with the evidence rather than a bare risk flag nobody trusts.
Meeting notes into structured records
Sellers do not update the CRM properly, and no amount of nagging changes that. An agent takes the call recording or transcript and the meeting invitation and produces the structured update: attendees mapped to contact records, competitors mentioned, budget and timeline statements captured verbatim, objections logged, next steps with owners and dates, and proposed stage or close date changes with the quoted evidence for each.
Two rules make this work. Proposed changes to commercially significant fields wait for the seller's confirmation, which takes seconds. And every written field carries a pointer back to the sentence it came from, so a manager reading the record can check it.
Where AI fails here
Bad data in stays bad data out. If your CRM is a wasteland of half-filled records, an agent will produce confident summaries of nothing. Pipeline inspection is only worth running where activity is actually logged. Fix the inputs first, or accept that the first phase of the work is the agent telling you how bad the inputs are, which is useful but is not the outcome you were sold.
Ambiguous ownership breaks more deployments than model quality does. When sales, finance and revenue operations each believe they own the forecast definition, or when nobody can say authoritatively what stage four means, an agent has nothing to enforce. Writing those rules down is a prerequisite, and it is often the hardest part of the project.
Be sceptical of pipeline probability scoring. Models trained on your closed deals will reproduce your history, including its biases, and they perform worst when the market changes, which is exactly when you need them. Use agents for evidence and hygiene, where they are reliable, before you use them for prediction, where they are not.
Commission disputes involve fairness, precedent and relationships. An agent should calculate, evidence and recommend. It should never issue the final word on someone's pay.
And change management is real. Sellers who feel inspected will stop logging activity, which starves the system. The framing that works is that the agent removes admin and defends their commission, not that it audits them. That is a leadership choice, not a technical one.
Where to start
Pick one queue with an unambiguous right answer and a short feedback loop. Quote validation and approval routing is usually the best first move: the rules exist, errors are measurable, and a human still approves everything external. CRM deduplication and enrichment is a strong second, because every later use case depends on it. Run in draft mode long enough to measure accuracy per case type, then promote the mechanical categories to autonomous handling.
If you want a straight answer on which part of your sales operations spine would benefit first, describe the process to Adopt AI Partners, including the parts we would not automate.