Agentic GTM: why data has to come before agents
In this article:
Agentic GTM is a way of running go-to-market where AI agents carry out work such as account research, outreach and CRM updates, while people set direction and review results. Agents act on whatever data they can reach, so outdated records lead to wrong actions at machine speed. GTM leaders name four needs: data agents can reach, context they can read, records that stay accurate, and clear controls over use.
At a glance
- Agents act on whatever data they can reach, so outdated or incomplete records get acted on at machine speed.
- Agent-ready data needs four things: access, context, accuracy and control.
- Start with the outcome you want and the data it depends on, then keep experimenting as the technology improves.
An AI agent spends 30 minutes researching a prospect, preparing a campaign and drafting the messaging. Then the team realises the person it targeted left the business 18 months ago.
That happened to Jocelyn Arcega, Head of Solutions Engineering at SnapLogic, as his team started running agents. He shared the story at Data Before Agents, a panel Cognism hosted at The Shard in London. Alongside him were Orian English, GTM Leader at OpenAI, and Viktoria Ruubel, CPTO at Cognism.
Their conversation was about agentic GTM, where AI agents research, prioritise and contact accounts while people take the helm. The underlying theme was that agents are only as useful as the data behind them, while organisations have four primary needs when it comes to implementing AI-driven workflows – access, context, accuracy and control.
What is agentic GTM?
Agentic GTM is a way of running go-to-market where AI agents carry out multi-step work. Examples include account research, outreach and CRM updates. People set direction and review the results, a chatbot answers a question, and an agent takes actions to reach a goal. Agentic sales applies the same idea to the day-to-day work of a sales team, from meeting preparation to follow-ups.
Jocelyn described how quickly this has moved. Two years ago, his team was building knowledge agents that answered questions, and they often failed. It took a year or more to gain a few percentage points of accuracy. Then the Model Context Protocol (MCP) arrived, and agents could connect to systems directly. His team moved on from retrieval and started connecting agents to the CRM and other applications.
The people building those agents have changed, too. Salespeople, customer success managers, and SDRs now build their own. Sales work is too varied for a fixed workflow, so these AI sales agents need to be fairly autonomous. Jocelyn's example is a rep with 30 minutes before a customer meeting. The agent pulls the interaction history and public information, then builds a presentation in the account's branding. Two years ago, that was unthinkable.
What does agentic GTM look like day to day?
At OpenAI, every account has its own agent. Orian explained that a dedicated GTM innovation team built a semantic layer that connects the company's systems of record. These include the CRM, data sources, and meeting transcripts. The same layer feeds the forecasting system, so everyone works from the same sources of truth.
For a rep, the day starts with a proactive nudge. It might arrive as a Slack message or a ping from Dots, OpenAI's always-on agent. The agent might say it picked something up in yesterday's call. It then asks whether to escalate it, update the CRM, or book a follow-up. Sales leaders can see where deals stand without chasing anyone, which makes coaching much easier.
Orian has watched the tool change during her time at OpenAI. It began as a Q&A chatbot in Slack, where a rep tagged it to get an answer. Today, it works proactively. Leaders can also vibe code a dashboard, knowing it reads the same data that RevOps and Finance use, so the whole business works from one consistent set of metrics.
Orian calls this setup the 'GTM brain' – an informal name rather than its official title. The hardest decision for the team was what to build each agent around. After watching how the organisation worked and talking to design partners, they chose the account. Next comes an organisation-level agent that tracks what a company is doing as a prospect or customer. After that comes an always-on agent for each individual.
Why does data matter more as agents take over?
Data matters more because an agent can only be as good as the information it works from. Viktoria emphasised that the agent is the vehicle that executes the workflow, but data and context decide the outcome. Better data gets a team to a better result – faster – and at lower cost.
The reverse is just as true. Run a team of agents on poor data and the problem scales with them. Viktoria's description was that "you're just multiplying the noise". The team spends time and resources, then sees no upside in conversion or revenue.
This is where AI data quality differs from the problem sales teams already know. A rep working from a poor list might notice something looks wrong and stop, but an agent doesn't. It carries on, and then does it again for the next 500 accounts. Good data quality for AI is what keeps that speed pointed in the right direction.
What happens when agents act on outdated data?
Agents act on outdated data without hesitation, and the cost raises its ugly head later down the line. Jocelyn remembers it vividly from earlier this year. His team had powerful agents that could search for a prospect and prepare the campaign and messaging. He described the moment it went wrong: "It runs for thirty minutes, crunches tokens, and then you realise at the end that the person it was targeting has left the business eighteen months ago."
His team caught it that time, although he noted that other teams might not, and could end up running outreach to people who no longer work at the company. The fix was to verify that information from the CRM is still up to date. The team also checks whether what the agent finds online is up to date, with LinkedIn being one example. Finally, they confirm the right profile, which he called “mission critical”.
Records go out of date every time someone changes role or company, yet agents run around the clock and don't know a record has gone stale – they simply act on it.
Viktoria raised a second risk: picture a swarm of agents doing outreach and working from data that isn't compliant. The problem multiplies by every customer they reach, including people on Do Not Call (DNC) lists. Outdated data wastes effort, and data used without the right controls then creates a risk that's much harder to undo.
What does an AI agent need from your data?
An AI agent needs data it can reach, understand, trust, and use within clear limits. Viktoria's framework for what makes data AI-ready has four parts: accessible, understandable, accurate and current, and controlled. Each one answers a different way an agent can fail.
| Need | What the agent needs | What goes wrong without it |
|---|---|---|
| Access | A route to the CRM, data sources and call transcripts | The agent works from a partial picture or guesses |
| Context | Definitions that explain what each field and metric means | The same question returns different answers |
| Accuracy | Records that are verified and refreshed continuously | Outreach goes to people who have left the business |
| Control | Rules on who and what can use which data | Agents reach people they shouldn't, or data outside their remit |
1. Accessible
Agents can only work with data they can reach.
Viktoria reiterated that your AI agent needs to be able to access the data. It's as simple as that. In practice, that means creating connections between the agent and your systems, which is the role MCP plays. Jocelyn's experience also shows why this matters. Once the team’s agents could connect to the CRM and other applications, they were able to draw on a much broader pool of data, and progress accelerated.
2. Understandable
Data needs context before an agent can do anything useful with it.
Orian explained why this matters at OpenAI: "You can ask a question of the data and get completely different answers depending on the context."
The team relies on RevOps to define which metrics matter, how they are measured, and which systems serve as the source of truth. Those shared definitions allow agents, dashboards and people to work from the same assumptions and arrive at consistent conclusions.
3. Accurate and current
Stale records aren't ready for agents.
Viktoria pointed out that AI is effectively a machine working 24 hours a day, acting on whatever information it finds. The data behind it needs continuous verification and refreshing so it remains accurate and current. A quarterly database clean-up cannot keep pace with an agent making decisions and taking actions every hour.
4. Controlled
Control covers both who can access data and what agents are allowed to do with it.
Orian described how OpenAI manages this through role-based access controls. A GTM leader can access datasets that direct reports cannot, while research datasets remain restricted to the relevant teams. Each role operates with its own permissions and access to the appropriate systems of record.
"You can ask a question of the data and get completely different answers depending on the context. And that’s why that central governed source is so important, and the quality of the data is so important."
Orian English, GTM Leader, OpenAI
If an agent can't access data that a user is authorised to view, it should make that limitation clear rather than inventing an answer.
Viktoria added that governance must extend beyond the data itself to the actions agents take. Compliance needs to be built in from the start because an agent operating at scale can repeat the same mistake thousands of times if the underlying controls are not in place.
When should agents act on a signal?
Agents should act when a signal says the timing is right, not as fast as they can. Jocelyn's view was that smart GTM is "about being precise more than having a quantity of activities". Buyers are overwhelmed by outreach, and a faster agent sends more of it. His team is moving from enrichment alone to signals, so the agent acts at the right moment. The fewer actions it takes, the better.
Viktoria explained that if you filter a list by location, industry, and job title, you may still end up with 500 people. An agent can work through all 500 faster than a rep, but speed doesn't produce the outcome. A signal does, though. That could be a hiring push, a company event, or someone moving jobs. And as Viktoria highlighted, it’s only "when the signal comes through, you act on it".
Sourcing signals is a separate challenge. Jocelyn pointed out that revenue teams could scrape the web for job adverts and LinkedIn posts themselves. It's painful, time-consuming work, and every team repeats the same effort. For many teams, buying that information is the more practical route.
Who owns agentic GTM?
The business now builds, which changes what RevOps and IT do. Viktoria believes every business user has become a builder because building is so easy. Reps, Finance, and product teams know the context and what matters – they only need access to the data.
That shifts the job for RevOps. Viktoria described them as "orchestrators of all the agents that have been built within the company": they make sure the right data is bought and supplied. IT then builds the connectors that let people use it. However, she noted that nobody knows what the RevOps role will look like in 12 months.
Orian agreed, with a caveat. OpenAI's business users vibe code dashboards and sites. The sales team now shares interactive sites with customers instead of slides, and customers can explore them afterwards. Even so, the company depends heavily on RevOps. They decide the metrics, the KPIs and the systems of record. Without that, every team could build something that gives a different answer.
Where should you start with agentic GTM?
The panel disagreed on where to start, and each answer has a case behind it. Jocelyn argued for focus. Modern technology lets business users create their own use cases, but a GTM organisation should still target high-impact ones. Pick the KPIs you want to move, such as revenue or pipeline, and let that decide where to spend effort. The technology is ready, so there's no reason to wait.
Orian took the opposite view. "The most important thing is just to get started", she said. Give the technology a new challenge. If it can't rise to it, try again a week later, because the model may have improved. She wouldn't worry about finding the perfect use case. What matters is that the organisation becomes AI literate and has an experimentation budget. Much of what gets built will be thrown away. In fact, every earlier version of OpenAI's GTM brain is already retired, and she doesn't think the current one would exist without those steps. Her advice is to plan a token budget that includes experimentation and to push teams to use the technology every day.
Viktoria said both. You need a specific outcome, so there's a roadmap that delivers a return. Spend will increasingly be judged against that return, whether it's measured in ROI or token cost. Technology is also changing fast, so teams need room to see what else is out there.
If you want a practical order of play, consider the following steps:
- Name one outcome the agent should move, such as meetings booked or pipeline created.
- List the data that the outcome depends on. Check it's reachable, understood, accurate and controlled.
- Start with one agent on one narrow task, and review what it produces before it acts.
- Set aside budget for experiments, and retry failed attempts as the models improve.
- Measure results against the outcome, then extend to the next task.
"You have a team of agents. You’re just multiplying the noise. You are not getting to the outcome you want to get to."
Viktoria Ruubel, CPTO, Cognism
What will agentic GTM look like in 12 months?
The panel closed with one-sentence predictions, and all three pointed at the same place. The strongest GTM teams will stop talking about the technology and how many agents they've built. They'll talk about outcomes, data quality, the right context, and intelligence. Agents will also become proactive and teach people to do their jobs better, from preparing presentations to planning next actions. Finally, nobody will chase their team for a CRM update again.
Each prediction depends on the same foundation – an agent can only coach a rep or update a record if the data it reads is accurate.
Get your data ready before you scale agents
Agents can already research, prioritise, and contact accounts at a pace no team can match, which makes the data underneath them the deciding factor. The panel's advice came down to naming the outcome, checking the data behind it, then letting agents run. Teams that get access, context, accuracy, and control right now will be ready for what comes next. Cognism helps by giving agents accurate B2B data to work from. See how Cognism MCP connects your agents to that data.
Frequently asked questions
Sales automation follows a fixed workflow. Each step is set in advance. Agentic GTM uses AI agents that work towards a goal and decide the steps themselves. They can take actions such as researching an account, updating the CRM, or drafting outreach, with people in the loop reviewing the results.
Agents need data they can reach, context that explains what the data means, records that stay accurate and current, and controls over who and what can use it. Typical sources include the CRM, meeting transcripts, and external data providers. Each source needs to meet all four needs.
They can, but they'll act on whatever the CRM contains. Outdated contacts, duplicates and missing fields lead to wrong actions at machine speed. Verifying CRM records before agents act on them is one of the most effective ways to prevent costly mistakes.
Business users such as reps and customer success managers build many of the agents, because they know the context. RevOps defines the metrics, systems of record and data supply, and orchestrates the agents across the company. IT builds the connectors between agents and data.
MCP is one way agents connect to systems such as a CRM, and it made connections much simpler for teams like SnapLogic's. Any approach works if it gives agents reliable access to current data. What matters is that the connection exists and the data behind it is accurate.
Start with one outcome, such as pipeline created, and identify the data it depends on. Run a small experiment with one agent, review what it does, then expand. Keep a budget for experiments, because models improve and earlier attempts are worth retrying.