Cookbook
Running a sales operation with an agentic CRM
Real workflows, not features. Each one is a job a team actually does, run against the real product before it was written down. If a job needs a step that tells you to work around something, it is not here yet.
Nothing matches that. The chapter titles are the other way in.
Set it up
The first week. Get what you already have into a workspace, shape the pipeline around how you sell, and model the things your business runs on.
- Populate an empty workspace from what you already have Load a messy export into a new workspace, and find out what your own data disagrees about. A populated CRM, deduplicated on identity, and a list of the fields your source contradicts itself on.
- Move over from HubSpot A migration with no importer, which sounds worse than it is. The hour spent deciding what to leave behind is the valuable one. Everything you need for Monday, in stages you chose, safe to re-run.
- Build a pipeline that matches how you actually sell Describe your stages in a sentence, and find out the one thing the product refuses. Your stages, your odds, and a pipeline in which a deal can actually close.
- Model something the CRM has never heard of Describe the thing your business actually runs on and get a real object type with typed, enforced fields. Your own type, with its own identifiers, usable exactly like a deal.
Run the motion
The daily and weekly work of selling, and the ninety seconds after each of them where the record either gets made or does not.
- Work an inbound lead from the form to a qualified deal The whole motion across five days, including the two places a traditional CRM would have used a rule and you should not. A record somebody else could pick up cold, and a deal with a real number.
- Brief yourself on an account in one question The five minutes before a call, done properly, including the part that tells you which of it somebody actually checked. A brief that names its own sources, so you know what to lead with and what to verify.
- Run a discovery call and leave the record correct Sixty seconds before and ninety after. The gap between them is where every CRM has quietly failed, and a recorder your agent already reads can close most of it. What you learned in sentences, every commitment a dated task, and a deal that says what you now believe.
- Run the Monday pipeline pass The weekly review as one conversation, so the decisions you make while reading it land before you close the tab. Stages moved, reasons recorded, follow-ups created, none of it left for later.
- Put a gate in front of a closing deal There is no workflow engine here, and that is the honest starting point. So the gate is the agent, and this is how to make one that actually holds. A stated close checklist the agent applies every time, and a record of what it found.
Look after the customers you have
What happens after somebody signs. The steps a CRM has no field for, the bad news that has to reach the account team, and the renewal that is decided months before anybody opens the contract.
- Run the same onboarding for every customer The steps after a deal closes are the ones a CRM never has a place for, so they end up in a spreadsheet nobody updates. Onboarding steps held per customer, askable by account, and countable while open.
- Keep an escalation where the account team will see it Support knows. The person walking into the renewal does not. Closing that gap is one decision about where the thing lives. An escalation on the record, and in the brief anybody reads before they call.
- Run the renewal motion Renewals are lost in the quiet months, not the negotiation. The monthly pass that finds the quiet before anybody is thinking about a contract. Every renewal a forecastable deal, and risks written down before the outcome is known.
Wire it to what you already use
Indraft holds the record; the tools around it do the moving. What belongs to an agent, what belongs to an automation, and how to connect either without writing code.
- Decide what belongs to an agent and what belongs to an automation Indraft has no workflow builder and will not grow one. Here is what replaces it, including the half that should not be an agent. A split you can defend, and the one thing that is genuinely gone.
- Keep your CRM, and put an agent beside it You are not replacing Salesforce this year. Run Indraft alongside it as the layer your agent writes to, and stop the nightly sync undoing what the agent learned. The incumbent still system of record, an agent writing freely beside it, and a report of every field the two disagree about.
- Run a recipe on a schedule, without being there Half the recipes here end with "run it every morning". This says where that schedule lives, what Claude and ChatGPT each do on a timer, and which jobs should not be on an agent's schedule at all. The reading jobs on your agent's own scheduler, the ones that would lose data on a tool that retries, and a way to tell whether either ran.
- Capture your customer email automatically Every CRM you have used had a mailbox sync. This one never will, and the outcome you actually wanted is better without it. Accounts that show what was said, a job safe to re-run, and no copy of your mail anywhere.
- Turn a form submission into work somebody owns Your CRM used to host the form. Wire the one you have into Make or n8n instead, and the two decisions in the middle are the actual recipe. A matched contact, their words kept verbatim, and a dated task nobody has to remember.
- Notify a channel when a deal moves The most-loved feature of every CRM and the first one you notice breaking. Two ways to build it without writing a handler, and why the simpler one is usually right. Notices in the channel, a public URL nobody else can use, and a failing endpoint that looks like one.
- Drive Indraft from n8n or Make, without writing code Two directions: Indraft telling your automation something happened, and your automation telling Indraft. Both are configuration. A working connection in both directions, and a delivery log you can read when one fails.
- Keep a spreadsheet or a warehouse in step Everything that changed since you last asked, with who changed it and how. The one read that makes an outside copy possible. A repeatable pull that never sends the same change twice and never misses one.
Trust what it says
Where a value came from, how to correct one so it stays corrected, and how to keep an agent inside the remit you gave it.
- Tell an agent's claim from a verified fact Before you repeat a number to a customer, ask which values nobody has checked and how the agent came by each one. Every value carries how it was come by, and a correction that survives the next import.
- Undo what an agent got wrong Take a change back without remembering what the value used to be, and learn the two cases where undo is the wrong tool. The earlier value restored, and the correction itself in the history.
- Give an agent only what it needs The first thing to do in a new workspace and the one people do last. An agent is never even shown a tool it cannot use. Each credential as narrow as its job, and a departing colleague's tokens weakened by demoting them.
Report on it
The number you have to defend, and an honest look at what you predicted last time.
- Forecast the quarter, and find out why you lose The number, computed over every deal rather than a page of them, and a countable answer to why the others got away. A pipeline total you can defend and a breakdown of lost reasons.
- Weight the pipeline, and check last month's call The exact call for weighted pipeline by owner by close month, and how to keep last month's number so you can find out whether it was right. One weighted number your BI can read on a schedule, and a record of what you called before you knew.
- Report the quarter by stage and by owner One question, answered by the database rather than by exporting to a spreadsheet. Including the row that has no name on it. Pipeline totalled by stage and by owner, and the unowned amount stated rather than hidden.
- Capture why you lost, in a form you can add up Everybody writes the reason in a note. Nobody can total the notes. This is the difference between a story and a number. Losses recorded against a fixed set of reasons, and a count per reason you can read.
- Prepare a quarterly review, both kinds One review is with your customer and one is about your own quarter. The second is where a CRM earns its keep or gets caught. Numbers that reconcile with last quarter, and an honest look at what you predicted.
Keep it clean
The housekeeping passes that stop a CRM decaying into something nobody believes. None of them takes long; skipping them is what costs.
- Find out what your CRM is missing Not what is wrong in it. What is absent from it, which is the part nobody looks at until a number has to be defended. A written list of what is unowned, unfilled and gone quiet, in the order worth fixing.
- Give every record an owner in one pass An unowned account is not on anybody's list. Fix the whole set at once, and find out what the batch refuses to do quietly. Every account and deal owned, and a report of anything the batch would not apply.
- Merge the duplicates, and refuse the one that looks identical Survey a workspace for records that are the same organisation, and learn why the strongest match on the page is the one to leave alone. One record per organisation with both histories attached, and the near-misses still separate.
- Retire a dead account without losing its history Get a finished customer out of your lists and your counts while keeping the three years of history that make it worth having. Out of every list, fully readable, and restorable if they come back.
- Hand over a book of accounts Move a leaver's accounts in one batch, change nothing else, and find the ownerless records nobody was watching. Every account owned, the handover attributed, and nothing else touched.
Businesses that are not a sales pipeline
Five operations whose central record is not a deal: an application that moves while the applicant stays, a candidate placed repeatedly, a gift that already happened, an asset under a site, a retainer measured by the month.
- Run an admissions funnel beside your admissions platform The cycle resets every year, the metric is yield, and one person can apply twice. None of that fits a sales pipeline. Both cycles comparable in one question, and a nightly check for deposits that never moved the stage.
- Run a recruitment desk where the candidate outlives the placement You will place the same person three times in five years. A record with one status on the human being cannot hold that. A candidate's whole history in one question, plus who is coming free and where the margin is.
- Run a fundraising programme, where the best report is about absence A gift is a completed fact, not a forecast, so there is no pipeline. The list that matters is who gave last year and has not given this year. Giving totalled by fund and appeal, and a lapsed-donor list with what it is worth.
- Run a field service operation with sites and assets, not just accounts Work happens at a site, to a specific machine. Two levels of hierarchy cannot say that, and three can. What is overdue and where a van has to go, plus what is down right now.
- Run a retainer business and see scope creep while you can still fix it Agencies do not lose money on deals they fail to win. They lose it on work they agreed to do for free, one favour at a time. Every month's overage visible in the month, with the sentence explaining each one.