How Personio built an AI-powered go-to-market in 6 months

What an AI-powered go-to-market actually looks like at Personio

Most writing about an AI-powered go-to-market is produced by companies selling the tools. This is not that. Philip Lacor is CRO at Personio, a Munich-headquartered HR and payroll platform with roughly 1,500 employees, 15,000 customers, and a sales organization of about 400 people. He spoke at SaaStr AI London about what his team actually built, what it cost, and what is still broken.

The timeline is the part that should make you uncomfortable. Personio ran an internal AI search week in May, kicked off by co-founder and CEO Hanno Renner, with speakers from OpenAI, Mistral, and AWS. Six weeks later Lacor opened a Slack channel called AI-powered go-to-market. Roughly six months after that, they were running 400 assistants in production.

Here is why the six-week gap matters. After AI week, 90% of the go-to-market team was using LLMs weekly. Lacor's read on that was blunt.

Although usage was high, this is maybe not enough to reach true transformation and to really fundamentally change the way we go to market.Philip Lacor, CRO, Personio

That instinct is backed by the data. McKinsey's State of AI survey of 1,993 respondents across 105 nations found that nearly two-thirds of organizations have not begun scaling AI beyond pilots, and that the roughly 6% who qualify as AI high performers are nearly three times as likely to have fundamentally redesigned individual workflows. Usage is table stakes. Workflow redesign is the variable.

Lesson 1: Bottom-up gets usage, top-down gets transformation

Personio did the bottom-up motion well. Everyone got LLM access, training, and a week of dedicated time to build.

But Lacor argues the bottom-up motion alone stalls at experimentation, and the reason is structural rather than cultural. Real workflow change requires decisions only leadership can make: resource allocation, budget, tool consolidation, and explicit permission for people to spend significant time changing how they work rather than hitting this quarter's number.

You need to give people permission to spend a lot of time on changing workflows and changing ways of working.Philip Lacor, CRO, Personio

That permission is the scarce resource, not enthusiasm. If your AI initiative is currently a Slack channel with no budget line and no named owner with authority, you have a bottom-up motion and you will stay in experimentation.

Lesson 2: The three-part team you cannot skip

Personio's AI-powered go-to-market working group started at 15 people, deliberately large, and kept growing. It has three constituencies, and Lacor is emphatic that removing any one of them breaks the output.

He has watched both failure modes at Personio. The data and systems team built things with LLMs that lacked business context, so the models performed poorly. Separately, salespeople wanted to build things and had no support from data, systems, or RevOps, so nothing shipped.

The connective role is the GTM engineer, a function Personio added only a few months before the talk. Lacor's spec is specific: business background, heavily data-driven, technology-focused. Not an engineer who learns sales, and not a salesperson who learns SQL. The mix is the job.

The working group was oversized on purpose. Fifteen people gives you coverage across every function, and more importantly it gives more people a hand in shaping the direction, which is what makes the culture change stick later.

Lesson 3: Prioritize with jobs to be done

This is the most copyable part of the whole talk.

Personio's problem was not a shortage of ideas. It was the opposite. Once the working group started shipping, ideas began flowing from everywhere, people raised their hands in Slack, and work started on new use cases before the first ones were finished. In Lacor's words, it started to spiral a little bit out of control.

First, jobs to be done, applied to roles rather than products

The question is what this role was hired to do. Personio ran it across SDR, sales, customer success, and solution engineering. One GTM engineer shadowed account managers for two weeks and produced a readout: they were working across seven or eight systems, and a single simple task required switching between them and reassembling context by hand. The quantified finding was that the job was losing about two and a half hours per day.

Second, map those jobs onto the customer journey

This is what turns a list of local time-savings into a coherent plan. It shows the team how the pieces connect and where the compounding effects are. Then pick based on pain, not novelty. Lacor's filter: where is your biggest growth problem, your biggest P&L challenge, or your biggest customer experience gap?

The two-week shadow is the step most teams skip. It is also the step that produces the number that unlocks the budget.

Lesson 4: Plan quality times acceptance

Lacor uses a formula for change management: the effect of a transformation equals the quality of the plan multiplied by acceptance. Five times five beats ten times one.

That framing is worth sitting with, because most AI go-to-market strategy work over-invests in the plan and under-invests in acceptance. Personio drove acceptance three ways.

• Lead it. In deal reviews, when an AE showed up with a PowerPoint about how they would win a deal, Lacor would stop them and walk them into Gong's AI account brief live. The PowerPoint became unnecessary in real time. He is explicit that his leads have to do this too, in the tools, teaching the team, not delegating it to enablement.
• Share it. Teams present what they built to their peers. One team built an assistant that personalizes customer decks. Another built one that answers RFPs. Peers demoing to peers moves adoption in a way that a rollout email does not.
• Celebrate it. Personio allocated two or three President's Club seats specifically for the best AI contributions, with more seats promised the following year. President's Club is the strongest status signal a sales org has. Spending it on AI contribution tells the team exactly what the company values now.

Lacor's number one hiring trait, before and after AI, is curiosity. The people who are very curious, who are really leaning in trying to figure out how this new world is working, are the ones who help you drive things forward.

Lesson 5: Your stack plus your context

Usually the tools are not the panacea. There is usually a lot of work that you need to do in your workflows, in your data, and other things. So let us start with what we have.Philip Lacor, CRO, Personio

The stack: Salesforce as CRM, Gong for conversational intelligence, Qualified for meeting booking and website intent, Snowflake as the data layer, and Amazon Bedrock as the LLM layer so they can swap models. Gong Engage replaced Groove for sequencing after Groove did not work out.

Two things they did before any of it worked well, and this is the unglamorous half of the story.

• They fixed the data. One third of their Salesforce records were duplicates, so they installed automatic deduplication. They spent months cleaning the prospect database and buying and connecting external data sources. That work is independent of AI and makes AI dramatically better.
• They loaded the context. About 5,000 Gong calls into Snowflake, plus emails and Salesforce data. Then the Personio-specific layer: ICP definitions, pitch decks, onboarding processes, and product training materials. Lacor calls this the critical step. Data volume alone does not produce a useful model. Company-specific context does.

The four use cases, with real numbers

Three observations that matter more than the numbers.

The expansion SDR case is the cleanest ROI story in the talk, and it is not really an AI story. It is a this person opens twenty systems a day story. The AI is just what finally made the integration cheap enough to build.

The intent model did not work at first. Lacor is direct that it was picking up signals they did not think were good, they changed it, and it improved. Budget for the tuning cycle.

Nia produces something Lacor did not expect and now reads daily: the chat logs. Prospects ask about minimum pricing at 11pm on a Friday. That transcript stream is unfiltered voice-of-customer data most companies would pay a research firm for.

What they still have not solved

This is the section the vendor guides leave out, and it is the most useful part of the talk.

• Context decay. Lacor raised it himself and did not pretend to have an answer. As they keep loading data, some of it goes stale, and stale context makes models worse rather than better. If you are building a GTM knowledge base right now, design the eviction policy before you need it.
• Agent-to-agent routing. Nia gets support questions. Their incoming Fin support agent will get commercial questions. How the agents hand off to each other is unresolved. Lacor guesses they land at three to five agents total.
• The four weeks of wasted demos. Nia went live before anyone owned training it. It started giving legal advice. It started bashing competitors. Demo requests are the best leads a company gets, and roughly four weeks passed before someone was assigned to train it daily. Assign the owner before launch, not after.
• The 400-assistant power law. Personio has 400 assistants and the top ten deliver roughly 80% of the value. That is not a criticism, it is a planning input. If you are choosing between building ten assistants properly and four hundred partially, the data says ten.
• The price tag. Asked whether Personio spends six figures on AI, Lacor said more. He put SDR agents at roughly $100K each. Almost no AI go-to-market case study includes a number like that, which makes it the most useful sentence in the talk for anyone building a business case.

One more unresolved question, and it is a good one. Personio tested rep coaching with Hyperbound. New hires will use it because onboarding forces them to. Will experienced reps voluntarily keep using an AI coach, or does usage tail off once the mandate ends? Nobody knows yet.

How to build your own AI-powered go-to-market

Lacor's closing advice, sequenced into something you can run this quarter.

1. Lead from the front, personally. Not a mandate. Open the tool in a deal review and show your reps what it does. If the CRO does not use it daily, nobody will.
2. Assemble the trio before the first build. Data and systems, a GTM engineer or RevOps person, and the business. One without the others produces work that does not ship or does not fit.
3. Shadow one role for two weeks. Pick the role closest to your biggest pain. Count the systems. Count the hours. That readout is your business case and your roadmap.
4. Map the jobs to your customer journey. This is what stops the idea backlog from spiraling and shows the team how the pieces compound.
5. Fix the data first. Deduplicate the CRM, clean the prospect database, load the context. Boring, unavoidable, and the reason most AI sales agents underperform.
6. Assign a named owner to every agent, before launch. Personio's Nia has one person accountable for reading the daily output, applying feedback, and testing in real time. That is the job, and it does not end.

The mistake Lacor most wants people to avoid is endlessly testing tools. You have got to dig in and go deep. It is about doing AI instead of learning AI. Pick two or three, go deep, accept that the training is the work.

And his honest framing of ROI, which is the right expectation to set with your board: it is not instant, and it shows up in more places than one. Deal velocity, pipeline quality, win rate, retention, and the fact that the automated work is usually the part of the job people liked least. His expansion SDRs use their assistant every day because it made the job better.