Pictet: a 50-directive compliance check took hours, not two weeks
Anthropic says Swiss bank Pictet ran a compliance gap analysis in hours that was sized at two weeks for three people. No cost or error rate is given.
At a glance
- Company
- Pictet · Financial services · Size not stated
- Team and use case
- Legal and compliance · Regulatory monitoring
- Tools
- Claude Code, Claude Cowork, Artefact (training partner)
- Stage
- Scaled across teams, since Beginning of 2026
- Results
- Gap analysis of 50+ internal directives: 2 weeks for 3 people (projected) → Hours (not quantified)
- Staff with Claude Code or Claude Cowork: About 700
- Staff trained in half-day workshops: More than 500, in 25 sessions
- Weekly AI newsletter: 3 hours a week → Automated
- Cost
- Not disclosed: no figure for licences, usage, the Artefact workshops or the API gateway build.
- People
- Not disclosed for the compliance team: the page does not say who checked the gap analysis or what the three-person team did instead. More than 500 staff took a half-day workshop.
- Evidence
- Vendor-reported
Published by the company selling the product, usually with the customer. Wins are selected; cost is usually missing.
What this leaves out: Anthropic published the page. The two-week baseline was a projection, not a measured run, and "hours" is not quantified. No error rate, reviewer, cost or date is given.
Pictet, the Swiss private bank, compared more than 50 internal directives against current regulatory standards "in just hours" using Claude Code, Anthropic's tool that carries out multi-step tasks on files and code. The job had been projected at two weeks for a dedicated three-person team. The figures are in an undated customer story on Anthropic's site, which quotes Steve Blanchet, Pictet's head of group technology strategy and innovation, and Xavier Meyer, its head of cloud engineering.
That is six person-weeks of planned work (3 people × 2 weeks). The baseline is a projection, not a measured run. The page does not say how many hours the analysis took, who checked the output, or how many gaps it found or missed.
How risk and compliance signed it off
- Three years of groundwork. Approval "took close work with the risk, information security, and data protection teams", a relationship the page says the group had been building for three years. The AI policy was built onto the bank's existing risk framework.
- Internal use only. Pictet aims AI at its own operational processes, not at client-facing products, which the page says helped.
- Data location. Business use cases must run in the EU, and some only in Switzerland, behind an API gateway the bank built for the purpose. Data can never persist for model training.
- One group at a time. "What also helps managing the relationship with risk and compliance is this rollout per persona and per kind of use case," Meyer says. Developers went first, because their work carries the loosest data-location limits.
The rollout in numbers
The rollout began at the start of 2026 in Pictet's 1,500-person technology division. About 700 people now have Claude Code or Claude Cowork, the version for staff outside engineering: mostly engineers, plus product and business colleagues. Artefact, a data and AI consultancy, ran a half-day workshop 25 times for groups of about 20 on Pictet's internal training platform. More than 500 people attended, and the first six to eight weeks booked out when the schedule went live.
Other results the page names: one team automated a weekly newsletter that took three hours a week. A hackathon tool, PicAccess, shows staff who change roles which application access their new colleagues hold and pre-fills the IT request for what is missing. The page gives the prototype figure two ways: "two hours instead of two weeks" in its summary, and "minutes instead of the one to two weeks a traditional cycle took" in the text.
Reading this at work? Get one story like it explained every morning.
A five-minute read every morning at 7am ET: one story explained, two in brief, what else happened, one number. Unsubscribe in one click.
What the page leaves out
- Cost. No figure for licences, usage, Artefact's fee or the gateway build.
- Checking. The page does not say who reviewed the gap analysis or how its findings were tested. It gives no error rate.
- The people. It does not say what the three-person team did instead, or whether any headcount changed.
- Dates. The page is undated and does not say when the compliance analysis ran.
What to do if you run compliance or IT at a regulated firm
- Compliance: pick one bounded comparison you have already sized, such as your policies against a new rule. Run it with an agent tool, have the team that would have done it check a sample by hand, and record the hours for both.
- IT: before buying seats, write the rollout order by group and data sensitivity. Get risk, security and data protection to agree which data may be processed in which country, and put the no-training term in the contract.
- Both: budget training with the licences. Pictet put more than 500 people through half-day sessions while giving about 700 access. "It's not only about rolling out a tool: it is about building capabilities and bringing people along," Blanchet says.
Take this to work
Forward this
Anthropic says Swiss bank Pictet used Claude Code to compare more than 50 internal directives against current regulations in hours, a job sized at two weeks for three people; the page gives no cost, error rate or reviewer. Before counting on a similar saving, run one comparison that has already been sized and have compliance check a sample by hand.
Questions to ask
- How many hours did the analysis take including review, and who signed off the findings?
- How many gaps did a human check find that the tool missed or flagged wrongly?
- What does a seat cost at our headcount once usage charges are added, and what did training cost per person?
- Which of our data may be processed outside our country, and will the contract say nothing is kept for model training?
The arithmetic
people the task was sized for × weeks it was sized at = person-weeks planned; subtract the hours spent running the tool and hand-checking its output
Pictet: 3 people × 2 weeks = 6 person-weeks planned. The page says the analysis took "hours" but does not give the number, the review time or the licence cost, so the net saving cannot be worked out from it.
Questions this article answers
What did Pictet use Claude Code for in compliance?
Anthropic's customer story says Pictet used Claude Code for a gap analysis comparing more than 50 internal directives against current regulatory standards. It says the work was completed in hours and had been projected at two weeks for a three-person team.
How many Pictet employees use Claude?
About 700 people have access to Claude Code or Claude Cowork, according to Anthropic's page: mostly engineers, plus product and business staff. The rollout began at the start of 2026 in the bank's 1,500-person technology division.
What did Pictet's Claude rollout cost?
Not disclosed. Anthropic's page gives no figure for licences, usage, the Artefact training workshops or the API gateway Pictet built.
Where does Pictet allow Claude to process its data?
The page says business use cases must run in the EU, and some only in Switzerland, behind a purpose-built API gateway, and that data can never persist for model training.
A CIO re-tested his 25 AI agents' results; two of three numbers weakened
A CIO put three headline results from 25 AI agents through five checks and one held. Operations leads can run the same checks before funding phase two.
Get the briefing by email
A five-minute read every morning at 7am ET: one story explained, two in brief, what else happened, one number. Unsubscribe in one click.