Skip to content
Build Marketing leader insights Optimise

AI in ABM: from experimentation to agentic scale in practice

Sapphire Smith
Sapphire Smith

This post represents anonymised insights from a marketing leaders roundtable discussion. All participant information has been removed to protect individual privacy while preserving valuable learnings for the broader Changemakers community.

 

Where ABM leaders are with AI

AI has moved firmly off the experimentation agenda and into the operational core of B2B marketing. For ABM practitioners this matters more than most: the precision ABM demands, from deep account intelligence to personalised content and timely activation, maps directly onto the tasks that AI can now accelerate at scale. The challenge is not whether to adopt it, but how to do so in a way that field teams will actually trust and use.

Most ABM leaders are still in pilot mode, experimenting with individual tools and use cases. A smaller number are scaling. It is possible, however, to go further. This session featured a presentation from a global IT infrastructure services company on how they are building a fully orchestrated, multi-agent AI system in-house with a team that started with no technical expertise. Their journey provides the practical backbone for the themes that follow, and shows where the real barriers between piloting and scale tend to sit.

Theme 1: Getting started with use cases

Map the work before building anything

AI adoption stalls when teams reach for technology before understanding the work it is supposed to improve. The more productive starting point is understanding how ABM work is currently done: where time is spent, where effort is manual, and where output is slow or inconsistent. Field marketers typically carry a heavy research and profiling burden, and understanding a large enterprise account well enough to run a strategic programme can take days. That is exactly the kind of concentrated, repeatable task where AI makes an immediate difference.

VP of ABM, global IT infrastructure services company: “We looked at each of these steps and understood from our field marketers what they spend their time doing. Then we started breaking that down at a granular level. If I’m looking at account research, how do we currently go about doing it? We mapped the as-is state, and we realised there was a lot of manual effort involved, so how do we use AI to augment that?”

This mapping exercise does two things: it produces a prioritised list of use cases where AI can genuinely reduce effort, and it sets a quality benchmark so teams can assess whether any given output is actually good enough to use.

Solve one use case at a time, then build on it

Once use cases are defined, the approach that works is iterative and narrow. Pick one, build a workflow or agent to solve it, test with real users on live work, gather feedback, and refine. Only once that use case is working reliably should the next one be tackled. An example is to begin with account research: refined over multiple rounds of testing with the people who rely on it, the time saving is significant.

Key output from implementation:
4–5 days → 15 mins account research time with a validated AI agent

Head of Global ABM Centre of Excellence, global IT infrastructure services company: “We started off rudimentary: a basic research of an account. Then we took it to the market, worked on live deals, live accounts, with the deal teams and field marketers, to get constant feedback. Is the information enough? Is the information correct? Each of these individual bots went through dozens of iterations to reach a point where the outcome was approved and vetted by the account partner who knows that account well.”

From individual bots to a single orchestrated agent

As individual use cases are validated, a new problem emerges: ten separate tools for ten separate tasks create their own friction. The natural next step is a single orchestrating agent that understands a query, calls the right capability, and returns a structured output, all through one conversational interface. That shift from a collection of bots to one orchestrated agent is what moves an AI programme from useful experiment to operational infrastructure.

Theme 2: Building agents you can trust

Put it in their language or lose the budget

Early agent outputs are unreliable. What turns that around is not better prompting in isolation, but sustained collaboration with the field experts who know what good looks like. Often it is a case of working through dozens of rounds of refinement, testing live on real accounts with deal teams, and only signing off outputs when the account partner confirms the intelligence is valid.

Head of Global ABM Centre of Excellence, global IT infrastructure services company: “We worked with actual people in the field who know those things very well. They act as a barometer to train the agent on what best or good quality looks like. We do the iterations again and again, keep telling it: no, this is not what I want, I want it like this. Then you get the next version, and work with your field experts until they say: yes, this looks like what I can use.”

Note: Microsoft Copilot was the main engine used.

Training on examples of what good looks like proves particularly effective for structured outputs. For example, one use case saw a bot trained on 20,000 historical RFPs, mapping internal and external examples to define a quality benchmark before the agent produced any live output.

Key output from implementation:
1 week → 50 mins proposal evaluation cycle with a trained RFP agent

Compliance is a gate, not an afterthought – design it in early

The most common risk areas are access to first-party contact data, personally identifiable information (PII) exposure, and publicly available information used in ways that may not meet data protection requirements. Just because data is publicly available does not mean there is legal permission to act on it – an industry assumption that is likely to attract regulatory scrutiny.

Head of Global ABM Centre of Excellence, global IT infrastructure services company: “We consciously made an effort: contact databases that we own, agents cannot access that. It is off-limits to them. Persona-related modules look at a person’s work commitments and public statements. No email addresses, no phone numbers, whether verified or not.”

The teams moving fastest through compliance review did the work early: documenting guardrails, limiting access to sensitive internal systems, and walking legal and CIO teams through the detail before seeking production approval. Where there is a top-down commitment to AI transformation, the commercial case rarely needs making. What earns production approval is the due diligence.

VP of ABM, global IT infrastructure services company: “We have never been asked to prove it. The mindset is there from top down. We have to be an AI-native marketing organisation. What we had to justify was to the CIO, in terms of compliance. We had to prove that we had done the due diligence: negligible hallucination, no data breach threat, no PII being exposed.”

Start with public data, layer in internal data once trust is proven

The first version of an agent does not need to connect to everything. Build on external, public web sources first and use that deployment to prove reliability before wiring in CRM, intent platforms and customer data. Whitelisting known, dependable sources reduces hallucination significantly at this stage.

Head of Global ABM Centre of Excellence, global IT infrastructure services company: “Getting the output in a validated, negligible hallucination state from external sources is the actual tough part. Once we solved that, we started adding the layers of internal data, which will enrich all the outputs.”

For those looking to generate actionable insight from intent, success comes from combining sources rather than relying on one signal. First-party website activity, seniority scoring of individuals showing intent, and third-party signals layered in a weighted scoring model produce far more reliable results than third-party intent data alone.

Theme 3: Managing team adoption effectively

Power users are your fastest route to adoption

Deploying AI agents is the easy part. Getting a distributed team to use them consistently and effectively is where most programmes run into difficulty. Rolling out to a limited beta cohort and tracking usage from the first two weeks reveals clear gradations: power users, moderate users, and those who have not engaged at all. Each needs a different approach, and power users, activated as internal advocates, drive adoption faster than any formal training programme.

Head of Global ABM Centre of Excellence, global IT infrastructure services company: “Someone who is a power user, we use them as our spokesperson across the community. People who are not very comfortable, we go on a much easier scale: let us handhold you into this. Start with three or four use cases. Are you comfortable? Now try something else. And it is a continuous process.”

It can also swing the other way when users start to create their own bots and workflows in silo. Agent proliferation, also known as ‘agent slop’, can then emerge as individuals build without visibility of what already exists. Active curation, monitoring and governance, as well as training existing agents rather than building new ones from scratch, keep the tool estate manageable and quality consistent.

Community drives more sustained engagement than training

When it comes to AI adoption, peer advocacy is a more powerful enablement mechanism than formal training. Building an internal community where people can share wins and challenges tends to drive more sustained engagement than top-down instruction (although this does help).

VP of ABM, global IT infrastructure services company: “Every week we have a community call for the field marketers doing ABM. That will consist of our team sharing the latest capabilities, or it will be one of those field marketers talking about a deal they just won, or a challenge they have. Getting that community working is a really good way of getting enablement out there. It is not people telling them what to do, it is enabling people to help themselves and each other.”

AI in ABM also challenges the traditional boundary between marketing and sales tooling. If you have forward-thinking commercial teams, share your tools with them rather than holding them back. This helps to identify further use cases and drive adoption.

Theme 4: Optimisation and ongoing governance

There is no end state – plan for continuous iteration from the start

AI deployments will never be ‘done’. New use cases emerge from unexpected users, new requirements surface once adoption broadens, and the technology continues to develop. Planning for continuous iteration from the outset, rather than treating any release as a finished product, changes how teams resource and govern their programmes.

Head of Global ABM Centre of Excellence, global IT infrastructure services company: “When it comes to AI, we will always be in the optimisation stage. You reach a point where you feel it is ready, it starts being used, and then you get new use cases, new fault lines showing up, which you need to correct and make better. I do not see an end state.”

In practice, governance means: access controls that respect individual permission levels, active curation of the agent estate to prevent proliferation, and regular feedback loops with field teams built into the operating rhythm, not treated as one-off project tasks.

The process is iterative

It’s important to note that going from scratch to a fully orchestrated agent takes time: 12 to 18 months to fully build. But the key thing is to start, and to start small. Solve for one use case and build out from there.

Key takeaways

1. Start with the work, not the technology

Map how ABM tasks are currently done, identify where manual effort is heaviest, and let use cases emerge from that. Account research is one of the most impactful starting points: with a validated agent, a process that takes days can be completed in under an hour.

2. Build one use case at a time and prove it before scaling

Research and preparation time can fall by up to 90%. Proposal evaluation cycles can drop from a week to under an hour. Content customisation time can reduce by around 80%. None of that happens without first proving reliability at the individual use case level.

3. Compliance is the bottleneck, not the business case

Where leadership has committed to AI transformation, the commercial argument is pre-made. The work involves getting agents through CIO and compliance approval: design guardrails early, keep first-party contact data off-limits in the first iteration, and walk stakeholders through the detail before seeking production sign-off.

4. Power users and community beat formal training every time

Segment beta users by engagement level from day one. Use power users as advocates. Build a peer community where wins and challenges are shared. This sustains adoption far more effectively than a training programme.

5. Optimisation is permanent – resource for it accordingly

Every release will surface new use cases and new failure points. The teams making the most progress treat iteration as an operating model, not a project phase.

The skills that matter are less about knowing which platforms to buy and more about curiosity, a willingness to experiment, and the discipline to involve field teams at every stage. The marketers building the most useful AI infrastructure are not technical specialists. They are people who understand the commercial context deeply enough to know what good output looks like, and are persistent enough to iterate until they get there.

From insight to action

Too often, AI is a solution looking for a problem. The teams that make it work start somewhere different: with the workflow, identifying where the manual effort is and where the real opportunities sit, then working out how AI can help.

If that framing resonates and you are looking for a partner to help you move from experimenting to building, Modern works with marketing leaders to do exactly that: from building a strategy and roadmap, through to auditing where the real AI opportunities sit in your current workflows and developing solutions your teams will use.

Want to be in the room when marketing leaders are having these discussions? Sign up the Changemakers community below:

Join the Changemakers community.

Share this post