From zero AI strategy to 22 agents in production in four months.

There was no AI strategy at Deffrennes. No budget line, nothing on a roadmap. Jackie heard someone on another podcast describe what they were doing with AI agents, could not stop thinking about it, and downloaded one a few days later to see for himself.

He is not a developer and has never written a line of code, which is why he had left the whole subject alone until these tools became something you could simply talk to. He connected Claude to the company’s Salesforce, which handles that kind of connection out of the box, and then ran into the obvious question: what do I actually do with this?

He went for lead generation, which is what most people try first. Over a weekend in May he produced 600 leads. They were useless, by his own account, and he spent as long deleting them as creating them, all of it in the live system. But by Sunday the output had got good enough to surprise him.

We started from nothing. There was no need for it. We tried something, realised how powerful it was, and then went off in every direction.

Jackie Desfrennes

That weekend only worked because of the three years before it. When Jackie took over, the company was still running on email and fax for work that systems should have been doing. Modernising it was the main thing he brought to the deal.

By the time the agents arrived, the processes had been rewritten, the ERP had moved to Odoo, the CRM to Salesforce, and everyone was on the Office suite. Odoo and Salesforce both support MCP, the standard agents use to reach business systems. The result was an IT setup the company genuinely understood, with very little technical debt.

This is the part that does not travel well in a podcast. An agent is only as useful as the systems it can reach, and Deffrennes had spent three years making its systems reachable. Run the same weekend at a company still tracking stock in a spreadsheet and nothing happens.

The first experiment made something clear quickly: generating more leads does not save anyone time, it makes more work. More leads means more calls, more quotes, more chasing. And the company had started the year short-staffed, with the stated aim of taking work off people.

So they sorted the projects into three kinds. Time saved: work that simply stops, like the 40% of emailed orders now entered automatically. Value created: work that did not exist before and that eats into the time the first kind freed up. Prospecting is the obvious example. And expertise brought in-house: jobs that used to go out to suppliers.

The third is where the surprise came. By the end of July the company had finished its 2026 and 2027 IT roadmap early and done more than was on it, because integrations budgeted across several years got completed in one go. Accounting and HR work went the same way.

The agents are all called Alfred. Every Wednesday there is an Alfred Café: a standing meeting with the team managers about what has been built, what it changed, and what to build next.

It went badly at first, and Jackie is blunt about why. He would turn up, show what he had automated, and watch people work out that at this rate there might not be a job left. That was never the point, and saying so once does not fix it. What fixed it was the format: the same meeting every week, with the people whose work was changing in the room.

The rollout followed the same logic. He started on his own. Two months later eight people had access, the managers plus a key user in each team, so every team could bring its own ideas back. The week after that, everyone had their own assistant. The Alfred Café is now turning into a governance session run team by team.

There is one thing in particular he wants to avoid: shadow IT. Jackie would rather the company run its own tools properly than have everyone quietly automating company systems from a personal ChatGPT account. He calls his version managed shadow IT. Bring the idea to the Alfred Café, and if it holds up, it gets built properly.

What I want is for everyone to have an assistant that helps them do their job better and look after our customers better. The customer stays at the centre of it.

Jackie Desfrennes

They prospect every day. They handle messages across several marketplaces. They reply to customers in languages nobody in the office speaks. They sort the shared inboxes for accounts, contact, sales and purchasing, then deal with what they find.

An order arriving by email is handled within 15 minutes: entered in Odoo, a quote request opened in Salesforce, and the right salesperson told what happens next. Two more agents work on the data itself, filling out customer records from 41 different sources and checking creditworthiness.

The quote agent is the one that shows the difference between automating something and exercising judgement. Asked to quote 5,000 units, it checks physical stock in Odoo, takes out what is already reserved, reports what is genuinely available and what is on its way, flags that the price requested sits below the list price, notes that this customer has no order history and so no discount, and asks whether to round to full cases rather than break one open. Before, that order would have gone through at 5,000 units, the warehouse would have opened a case, repacked 20 units and lost three minutes. Once, that is nothing. Repeated all year, that is the margin.

The dashboards came with the agents. Jackie asks for one in plain language and gets it. What he wants from a dashboard is not a number but something to act on: a monthly reading and a recommendation, so the thinking is already done when he opens it.

The standard warning about AI is garbage in, garbage out. Deffrennes ran straight into it, and it turned out to be the best thing that happened.

The data was not so much wrong as old and full of duplicates. Putting agents to work on it forced the company to look at it properly and put some rules around it. Of 230,000 account and contact records, close to 100,000 were deleted: duplicates, records too poor to use, and companies Deffrennes does not trade with.

An agent checking the website found around 300 pricing problems in its first week, across 1,300 products. Volume discounts running the wrong way. Prices below what the product cost to buy. Prices out of step with the market, or with near-identical products sitting next to them. Nobody is to blame for that. It is what happens when no one has time to audit a catalogue and small errors pile up over years.

The bigger gain is what Jackie calls weak signals: things the company could never measure because measuring them was never worth someone’s salary. One region where a particular carrier is always late, so the cheapest option turns out to be the wrong one. A 15-person company cannot put an analyst on that. An agent can.

The main return for me is the data quality we’ve gained, and the service we can give customers because of it. We’re simply better at it now.

Jackie Desfrennes

On the quarterly dashboard, the central agent handled 22% of quotes, 45% of orders and 27% of claims. Those numbers understate it, because everyone now has an assistant creating quotes, orders and claims of their own. Counting those, quotes and orders are both near 70%, and claims around 90%.

Behind those percentages: 1,500 emails sorted over the quarter, producing 550 orders, 36 quotes and 114 claims, alongside automatic follow-up on abandoned baskets and replies on the marketplaces.

A quote used to take a day and a half to two days. The request landed in a shared inbox checked twice a day, went to the sales director, then to a salesperson, and got dealt with the next day if they had time. It now takes 15 minutes, at any hour, and it reaches the salesperson with the customer’s history, their credit check, the discounts they qualify for and a suggested approach already attached.

An order could take 48 hours to enter in the busiest week of the year. It now reaches the warehouse in five minutes, and the average time from an emailed order to delivery is just under three days. That end-to-end figure did not exist before, because no single system could see the whole chain.

In money: several hundred thousand euros over about three quarters, from April onwards. Jackie is careful about how he reads that. The 90 days a quarter are not 90 days of salary saved, because nobody left. They are 90 days of capacity the company did not have to go out and buy.

If I give people five hours back a day, they’re not going to spend five hours on leads. They’ll get maybe two hours free, and they’ll use the other three to do what they were already doing, better.

Jackie Desfrennes

The day before this conversation, Jackie deleted all 22 agents. Claude would not start, the automations were stuck, he asked what to do and approved the answer without reading it properly. The answer was to delete everything that had been built so it could start clean. Four months of work, one click. There were backups and it was all recovered quickly. He calls it human stupidity, and the mechanism is worth naming: long answers, read too fast, approved on reflex.

The one they still have not cracked is technical drawings. Deffrennes has its own design studio, and generating product plans automatically would change a lot. Claude is not strong on graphics, so they chain models together: Claude agents writing prompts that call the ChatGPT API to produce visuals that go back into other tools. It is still not good enough. They are still working on it.

And it costs time. For several weeks Jackie was going to bed at half past two in the morning. Moving this fast is one of the real advantages, but somebody has to put the hours in.

Start with one simple use case and accept that it will be bad. The lead experiment produced 600 useless leads, and that was the useful part: it taught him to ask for leads that look like his actual customers, then for larger accounts rather than every company in the country.

Get the right people around you. Someone who knows the business processes and someone who knows the IT, because the AI expert does not really exist yet. If you have neither, that is what an outside partner is for.

The cost is not the obstacle people imagine. A basic subscription is 15 to 20 euros a month, and the usage limits stop the spend running away rather than letting it.

And if he were doing it again: write far more of it down, and go slower. Fast returns are one of the genuine strengths of this technology, and also how you end up with 22 agents nobody has documented. Bring in the people whose work is changing as early as you can. They know the job, and they are the ones who will use it.

I’d document a lot more and go slower. And I’d bring in the people affected much earlier, because they’re the ones with the expertise and they’ll be the first to use it.

Jackie Desfrennes

• Sort the systems out first. Three years of moving to Odoo and Salesforce is what made that weekend possible.
• Separate time saved from value created. Only one of the two actually takes work off your team.
• Put a standing weekly meeting in the diary. The technology was never the hard part. The worry in the room was.
• Expect cleaner data to be worth more than the hours saved. That is the result Jackie rates highest, and nobody planned it.
• Read what the agent is asking you before you click approve.