← All Blog Posts

The toil is the project: what a security review taught us about AI project management

November 19, 2026
By Keshav Puttaaswamy

Most of the projects our customers run in Empwr are planned ones. They are longer efforts that take a fair amount of coordination, and they run anywhere from a couple of months to a year or more. We work the same way inside Empwr: every one of our ongoing product engineering projects runs in Empwr.

Recently we found out that Empwr is just as useful for the other kind of project: the unplanned spike that lands with no warning and has to be finished in a week. Ours was a security review.

The email that turned into a project

One Friday, a prospect in healthcare sent us a 26-page document of follow-up questions. We had already completed their vendor security questionnaire a few weeks earlier. This was round two, and it was bigger: 60 questions across six sections, covering data residency, AI and model governance, how patient data could end up in our product, contract terms and compliance controls.

Twenty of the 60 were marked critical. The reviewer couldn't continue the review until we had answered all of them. The bar was high for a good reason. Healthcare organizations answer to HIPAA, so every vendor that might touch their data gets examined closely.

Each answer had to be accurate, because a commitment in a security response becomes a commitment in the contract. Many answers also needed evidence: a document, a diagram, a certificate or a signed agreement, attached and named.

The work fell to four or five of us, across two time zones. Ludovic Hauduc, our CEO, and Chris Pirich, our CTO, owned most of the technical and legal answers. I owned the response and the relationship. Chris summed up the job in one word: a "root canal."

Our biggest challenge wasn't any single question. It was capacity. None of us is a project manager, and none of us had room to become one for a couple of weeks. Everyone's plate was already full: that same stretch we were shipping product, running customer calls and preparing for a board meeting. Someone needed to track 60 answers and the evidence behind them, chase owners and know when we were done. There was nobody to hand that job to. That is the gap Empwr filled.

The hard part wasn't the answers

AI made drafting the answers the easy part. Chris's view was that most of them could be worked out from our code and our existing documents, with only a few needing a human decision. He was right. Within a few days we had a full draft.

The hard part was everything around the answers. When the four of us met the following Monday to agree on an approach, most of the 45 minutes went to questions like these:

  • Which version is current? Two people had each started a draft of the response. Merging them risked overwriting someone's work, so we had to pick one.
  • Which document do we send? We had several data processing agreements, each written for a past customer. Nobody was sure which one should be the starting point.
  • Where does it live? Ludo put it plainly: finding a contract should take three clicks, and it was taking five messages to different people.
  • What's still hiding in the draft? Notes like "verify this" had ended up at the end of spreadsheet cells, where nobody had seen them. Each one needed checking, then removing before anything went out.

None of this is unusual. Any team answering a long security review hits the same problems. A project manager would normally absorb them: tracking who owns each answer, chasing evidence, keeping one version of the truth and knowing when everything is done. We didn't have one, and nobody had the capacity to step into the role. Without help, that work would have landed on all of us, on top of the answers themselves.

We ran it as a project in Empwr

We already had an Empwr project for this prospect. Empwr had been keeping it current since our first demo in June, from our meetings and email threads with their team. So that's where the security response went too.

That choice mattered more than it might seem. I could have tracked all of this in a Claude doc of my own, or even shared that doc with the team. But that would have created one more silo: a separate copy of the plan, cut off from everything else we knew about this prospect. The Empwr project was already shared by everyone working with this prospect, and it sat alongside every other project we run. Putting the security response there gave all of us the same context, without any setup and without anyone typing tasks into a tracker.

At the end of that week, we broke everything left to do into 26 tasks in that Empwr project. Claude created them for us from our team's Slack channel, straight into Empwr. Each task was written to be picked up by someone who hadn't been in the meeting:

  • The title leads with the question numbers it answers, for example "C13: Set the incident notification window." Anyone could see at a glance which part of the questionnaire a task closed.
  • The notes say why it matters and what done looks like. For example, one note said that five of our answers claimed documentation didn't exist when it now did, and that four of those questions were critical.
  • Every task has one owner. Chris took the questions about our AI providers, data flows and incident response. Ludo took the data processing agreement, the insurance certificate and the policy updates. I took the response document itself and the final send.

We gave ourselves one week to finish: start on Monday, send by the end of the day on Friday. That was the whole setup. Nobody wrote a project plan in a spreadsheet or a private doc, nobody entered tasks by hand, and nobody had to keep a separate tracker up to date.

Empwr became the place the work lived

By Tuesday, the Empwr tasks were doing more than tracking the work. They were holding it.

When Chris finished a question, Claude wrote the answer into the task itself. His Slack messages became short: "EMP-1338 is marked complete with the answers to question D8 within the task text." I now had the task to copy the answer into the response document. Empwr ensured the hand-offs didn't fall through the cracks. The task recorded who answered, what they said and when. Nobody had to search a long Slack thread for the latest wording. And because the context was shared, it didn't matter who was looking or where they looked from. Chris, Ludo and I saw the same tasks, owners and answers, whether we opened Empwr directly or asked Claude about it in Slack.

The Empwr project also told us where we stood without anyone writing a status update. Ludo said in our weekly team meeting that Monday that he was loving using Empwr to manage the project. My rule for the team was simple: every task in the Empwr project gets checked off before we send. By Thursday afternoon, that rule had a precise answer: two tasks left, both waiting on our CISO.

The moment I remember best came the same afternoon. I noticed that Ludo had an open question with our CISO about the list of subprocessors in our data processing agreement, and that no task was tracking it. I said so in our Slack channel, and a minute later a new task was in Empwr, assigned to Ludo. Its notes held the open question, the decisions we had already made, links to the draft and the Slack thread, and what done meant: our CISO answers, and the list matches the one on our public Trust Center. Ludo closed it the next day. When a conversation turns up new work, Empwr suggests a task in the right project, with an owner, before anyone has to ask.

That is the pattern I want from AI project management. If work exists, it is tracked. If it is tracked, it has an owner and a definition of done. And keeping it that way doesn't depend on anyone remembering to do it.

Claude checked the work, and Empwr made sure the fixes got done

We used Claude throughout, to draft answers, check them against the questions and review our documents. Claude finds problems quickly. What usually goes wrong is the step after: a problem gets flagged in a chat, everyone agrees it should be fixed, and then nobody fixes it. With Empwr, a problem Claude found went onto someone's task list, or into the task already tracking that answer, so it got fixed. Three examples:

  • The wrong insurance document. The questionnaire asked for a certificate of insurance: a one-page summary of carrier, limits and policy period. Ludo had uploaded our two full insurance policies. Claude pointed out that neither was the certificate we'd been asked for. The fix went on Ludo's task list, and on Monday he got the one-page certificate from our insurer.
  • Answers that undersold us. Five answers about our Jira integration still said the documentation didn't exist. By then it did: Chris had written a detailed document covering exactly which Jira data we read, the access we request and how long we keep it. Four of those five questions were critical. One task got all five rewritten.
  • Contracts that had drifted. While Ludo was finishing our data processing agreement, Claude noticed our standard subscription agreement no longer matched it in places. Separately, Claude caught a typo in that agreement's notice email address. Both were fixed before either document went out.

None of these would have stopped the response from going out. Each one would have cost us credibility with a reviewer who reads every line, at the stage where trust is the whole point.

Done by Friday

On Friday afternoon, I did a final pass (three passes, really) and hit send. The package held:

  • answers to all 60 follow-up questions, with the vendor attestation signed
  • a revised version of the original vendor questionnaire
  • 14 supporting documents, including our SOC 2 Type 2 report
  • a document manifest listing every file by name and the questions it answers, so the reviewer could check our evidence instead of hunting for it

That was one working week, Monday to Friday, from kicking off the tasks in Empwr to sending the response, and every task in the Empwr project was done. Remember that none of us worked on this full time. Each of us fitted our tasks in between customer calls, product work and the rest of the day job. We finished in a week largely because Empwr did the project management for us. Each of us could see what was ours and what was left, and nobody had to stop and rebuild the plan to find out.

The payoff came early. The day before we sent it, another prospect asked us a set of security questions. Ludo answered them in about 20 minutes, using the documents we had just pulled together. In his words, it "would have been a much heavier lift in isolation." The work we did for one review is now ready for the next one.

What AI project management can be

Looking back, the split between the two tools was clear. Claude did the thinking work: drafting answers, comparing them with our documents and spotting what didn't line up. Empwr did the work a project manager would otherwise carry. It knew what needed to be done, who owned each piece and when it was due, and it kept that picture current as the work moved.

People often ask us why they need Empwr if they already have Claude. This project is my answer. Claude is excellent inside a conversation, but what it produces tends to stay there: a doc in one person's chat, a list only its author keeps current. Empwr gives the whole team one shared, always-current picture of the project, sitting alongside every other project we run. Claude can read from that picture and write back to it, and so can everyone else on the team. Claude made us faster at the answers. Empwr is why those answers came together, from a team with full plates, by Friday. Claude plus Empwr was far better than Claude alone.

That second kind of work is easy to underestimate. Updating a tracker, chasing an owner, rebuilding the latest status from a long Slack thread, asking "is this tracked anywhere?" Each of these takes a few minutes. Across a team, they add up to hours a week that do nothing to get the job itself done. That's the toil. Empwr took most of it off our plates, so the time went into the answers and the evidence instead.

We built Empwr with longer, planned projects in mind, and that is still where most of our customers use it. This week showed us it is just as valuable when an unplanned spike lands on a team with no spare capacity. We're still scratching the surface of what AI project management can do for us. The next security questionnaire is already easier, and the one after that will be easier still.

If you want to see how this works on one of your own projects, you can start for free or book a demo.

Ready to get started?

See the results for yourself.

More Blog Posts