45 deals sat in a stage called Closed. Nobody could say which ones were won.
Konect DCS sells data centre power and cooling equipment for partner factories and earns commission on each order, with Ampvik as the new brand beside it. The founder had built 21 workflows himself, and maintaining them had become the job. He wanted a Claude agent to tell him what to do each morning. The audit showed why it could not work yet: nothing in the workspace was dated. We fixed the data first, then built the agents and a QuoteWerks sync on top of it.
Free audit + paid build
Engagement
15 days
Audit to handover
21 to 12
Workflows
45
Closed deals split won/lost
72%
Win rate, first time measurable
~12 seconds
Quote sync round trip
2
AI agents
The problem.
The founder had moved off HubSpot nine months earlier and built the Attio workspace himself: four objects, a customer success, verification and escalation loop, a commission engine, and 21 workflows. It worked well enough that nothing looked broken, so nothing got fixed. On a sales cycle of six to nine months, from first conversation to a factory shipping the order, that is a long time for small errors to pile up.
The pipeline could not say whether a deal was won or lost. 45 deals sat in a stage called Closed, next to a separate Closed Won. There was no Closed Lost, no loss reason, and no expected close date anywhere in the schema. No win rate, no forecast, no cycle time.
Commitments lived in a text box. Next Step Due was a text field holding real deadlines like "Lines 5&6 delayed to 17-Oct", in a format nothing could sort or alert on. Only 21 of 122 deals had one, and all 7 open tasks in the workspace were past their deadline.
The automations had drifted out from under him. All 21 workflows were on Attio's legacy builder, three had unpublished edits sitting in the editor, and the escalation loop only closed when someone pressed a button. A save on a commission record failed silently because five fields were required, including payments that had not happened yet. And a workflow meant to score fit had flagged 50 of 169 companies as Do Not Target, including the group that owns his main factory partner.
Objectives.
- Make the pipeline record won and lost, with a reason, so a win rate exists at all.
- Turn next steps into real dates, because a prioritisation agent has nothing to rank without them.
- Cut the workflow count and move arithmetic into formula attributes that cannot drift.
- Model two trading entities, Konect DCS and Ampvik, in one workspace instead of renaming one into the other.
- Stop retyping deals into QuoteWerks, the quoting tool that stays the source of truth for quotes.
- Build the agent he actually asked for: a morning list of what needs him today, running on his own Claude account.
What the audit found.
A free audit over 48 working hours with read-only access, delivered as one document. Every finding checked against the live workspace. The four that shaped the build:
The agent could not be built first
An agent reading the workspace that morning would find prose next steps, no due dates, no close dates and seven overdue tasks. It would rank everything by nothing. Fixing the pipeline and the dates was not a warm-up. It was the agent's input data.
The silent save failure
Commission records had five required fields, three of them future payments. A record could not save until money that had not arrived was typed in. Attio showed no error, so the stage move just did not happen.
Workflows doing a formula's job
Six of the 21 workflows existed only to do arithmetic on commission and margin. The sums were correct to the cent. The problem was upkeep, and the lag: a workflow recalculates a beat later, a formula the moment a number changes.
No field for the legal entity
Two companies trade out of one workspace, with different quote templates and different accounting systems. Nothing recorded which one owned a deal, so no integration could know where to send it.
The rebuild.
A full backup came first: every object, every list, the whole schema. Record counts were identical before and after, and no record was deleted.
A pipeline that ends
Closed Lost, eight loss reasons, an expected PO date and a Priority field added. The founder split the 45 Closed deals himself: 73 Closed Won and 29 Closed Lost across the board, and a 72% win rate that could be measured for the first time.
Dated next steps
A real Next step due date beside the old text, which was kept and renamed Next step. Every morning Attio turns an overdue date into a task, and the agent reads the same field.
21 workflows down to 12
Seven deleted, six replaced by two formula attributes, the escalation loop rebuilt to move on a status change, and four new ones, including one that stamps the date a claim was first sent to the manufacturer and keeps it through a rejection.
Two entities, one workspace
An Owning entity field on every deal and commission record. It decides the quote template and which accounting system gets the invoice.
A commission pipeline that can move
Ten stages cut to six, the three payment fields made optional, and the manufacturer moved off a dropdown onto real company records, so the system knows who to bill.
Labels that mean one thing
33 partner and supplier companies wrongly flagged Do Not Target restored the first day, the rest resolved the next. Do Not Target now means never contact, not low quality. Garbled API field names were renamed so an integration can read them.
QuoteWerks and the agents.
The founder's own scope put the Claude agent at the centre. Everything else in the build existed so it had something true to read.
Attio to QuoteWerks
Tick Send to QuoteWerks on a deal and the quote document appears in QuoteWerks with customer, contact, addresses and currency filled in, in about twelve seconds. QuoteWerks mints the quote number and writes it back. It runs inside Attio, with no server to keep alive.
QuoteWerks back to Attio
Every hour a job reads QuoteWerks and updates each linked deal: total, revision, status, expiry, sales order number. It never moves a deal stage. A quote turning into an order is a hint, not proof.
Daily focus agent
08:00 Oslo time on weekdays it names the few deals that need him today, ranked by money at risk, and writes the list to the top of a page in his notes app. A second run at 15:00 on Fridays reports what moved and what went quiet.
Sales assist agent
Four jobs on demand: brief me before this meeting, summarise this deal, draft a follow-up, check the workspace for broken data. It drafts email and never sends it.
On his own account
Both agents run on the founder's own Claude subscription, from code in his own GitHub organisation. They keep working after our access is removed, with no separate AI bill.
Call recording on
No meeting in the workspace had a transcript. Attio's recorder now captures external calls, so the agents stop being deaf to what was said.
What we did not automate, on purpose.
A rule says deals in Norway belong to Ampvik and deals elsewhere to Konect DCS. It is a good rule, but the destination country was filled on 21 of 125 deals. A workflow reading it would have classified 17% of the pipeline and left the rest untouched, which is worse than manual because it looks automatic. The same thinking kept line items and part numbers out of the QuoteWerks sync, kept invoicing to drafts a human approves, and left accounting switched off: the Tripletex leg is tested end to end in its sandbox, and Xero waits on production credentials and a VAT answer from the accountant.
Results.
- A pipeline that records won and lost with a reason, and a win rate the business had never been able to compute.
- 21 legacy workflows replaced by 12, with the commission maths moved into formulas that recalculate instantly.
- Konect DCS and Ampvik modelled side by side in one workspace, with every deal and commission record owned by one of them.
- Deals push to QuoteWerks in about twelve seconds, and quote status, totals and sales order numbers come back every hour.
- Quote numbers pulled out of deal names into a real field on 54 deals. Eleven were left alone because they shared a number, and guessing would have linked the wrong quote.
- A morning focus list and a Friday review, running on the founder's own Claude account and landing where he already works.
- A runbook and a field-by-field data dictionary handed over, so the next hire can learn the system from the page.
Ready when you are.
Book a call and we will tell you honestly whether this is worth doing, or start with the free 48-hour audit and decide afterwards.