Event debrief

TBM Summit EMEA 2026: what the experience told us about the state of agentic AI in TBM

TBM Summit EMEA this year landed on the same problem from every angle. A panel discussion pulling together voices from across the vendor and practitioner landscape. A working session on AI productivity inside the TBM office itself. Hallway conversations in between. Different formats, different speakers, same conclusion.

TBM teams do not have a data problem anymore.
They have an action problem.

The gap is not visibility, it is execution

The panel was blunt about this. One point landed directly: the market has plenty of platforms generating recommendations, sometimes thousands of them, but far fewer ways to turn those recommendations into engineering action. Cost visibility exists. What is missing is the connective tissue between a flagged anomaly and a resolved ticket.

The AI productivity session reached the same place from the practitioner side. TBM teams spend a large share of their time on data quality reviews, anomaly investigation, and report building. That workload is exactly what keeps practice leads out of the value conversations they are supposed to be having with the business. The tools generate insight. The people generate the follow-through. And follow-through is where the hours go.

This is the real argument for agentic AI in TBM. Not smarter dashboards. Fewer manual hops between insight and outcome.

Agents augment, they do not replace

Every speaker who talked about agentic AI drew the same line: agents extend human judgment, they do not substitute for it. Archaius’s presentation described deploying agents selectively while keeping humans responsible for strategic intent and domain context. Yarken’s session made the same point from a product angle, framing agentic AI as a way to democratize TBM data and accelerate adoption, with guardrails that keep a human in control of the decision.

The AI productivity session unpacked why this distinction matters in practice. An agentic system is not one model doing everything. It is an LLM interpreting a request, handing it to a specialized agent such as a variance agent or a budget agent, with defined tools, memory, and guardrails shaping what that agent is allowed to touch. That structure is what makes delegation safe. You are not asking one general-purpose model to run your cost allocation. You are asking a bounded, auditable worker to do one job well.

The framing that came up more than once: treat the agent like you would train a new employee. Define the process the way you would for a person. Then translate it into instructions the agent can follow and be held to.

Trust is built through traceability, not promises

If there was one word that tied the whole summit together, it was governance. Yarken’s session on the panel emphasized secure, row-level governance and SOC 2 compliance as prerequisites for putting agentic AI anywhere near financial data. The AI productivity session went further into what earns trust day to day: the agent has to show its steps, cite the data it queried, and confirm it stayed within its guardrails. That is what prevents the black box problem and what stops hallucinated numbers from making it into a board deck.

None of that comes from a friendlier interface. Trust in an agent comes from embedding the right context, the right data, the right tools, and the right skills into a repeatable, standard operating procedure that the agent is bound to follow every time, the same way you would document a process for a new hire. That is a materially different approach from bolting a chatbot onto an existing report or dashboard, which is what a large share of the market is currently shipping. A chatbot answers a question. A governed agent, embedded in a documented playbook, actually does the work and can show its receipts.

This is not a compliance checkbox. It is the actual mechanism that lets a TBM lead hand off a variance reconciliation or an MSP invoice audit without staying up at night. Auditability is the product, as much as the automation is.

Vendor decisions need the same continuous model as cost data

Archaius’s session on the panel applied this same logic to vendor governance. Software and vendor decisions have historically been point-in-time events: evaluate, sign, forget until renewal. The argument was that this model is already broken, because market conditions, pricing, and vendor terms shift faster than annual renewal cycles account for. The proposed fix reads a lot like what TBM has been arguing about cost data for years: combine internal system-of-record data with external market intelligence, and treat the decision as continuous rather than a one-time event.

Put those two threads together, cost visibility that needs continuous encoding into action, and vendor decisions that need continuous market context, and a pattern comes into focus. TBM is moving away from periodic reporting toward standing, monitored systems. That shift is what makes agentic AI relevant here in the first place. Agents do not get tired of checking the same thing every week. People do.

A shared data layer is the foundation, not a feature

Archaius’s presentation of a common ontology, linking strategy, portfolio finance, TBM and FinOps data, and delivery cost, was one of the more structural points raised. Without that shared layer, agents have nothing consistent to reason over, and different teams keep answering the same business question with different numbers.

This lines up directly with what came out of the AI productivity session about model complexity. TBM models tend to grow into what one participant called spaghetti: too many inputs, too much custom logic, no clear line back to the business question they were built to answer. The advice was not to have an agent build the whole model. It was to use agents to accelerate the data analysis and ingestion work, so teams can justify discarding the overcomplicated model and rebuild something simpler and better documented.

A semantic layer is what makes both of those things possible at once. It gives agents something reliable to act on, and it gives humans a plain-language way to ask questions of financial data without needing to know where every number lives.

Where to actually start

The overall message was realistic about the fact that most TBM offices are starting from zero on this. The advice converged on the same starting point: contracts.

Contract management came up again and again as the first practical use case. Contract data is usually incomplete, scattered across systems, and manually chased down before every renewal. It is a bounded, well-understood problem, which makes it a good first target for an agent, and a good place to build organizational trust in agentic AI before extending it into anything higher stakes like forecasting or P&L interpretation.

The other piece of practical advice from the AI productivity session: stop chasing perfect data before you start. Define what “good enough” looks like for the decision at hand, and build from there. Waiting for a clean CMDB before deploying any automation is a good way to never deploy automation.

The TBM Analyst role is changing shape, not disappearing

The summit also touched on what this means for the people doing the work. Post-panel conversation traced how far TBM has come as a recognized discipline, and how hard it has always been to find people who combine IT, finance, and business fluency in one person. The AI productivity session picked up that thread with a specific prediction: as agents take on more of the data processing and report generation work, the TBM Analyst role shifts toward an embedded partner who helps business leaders interpret what the AI is surfacing.

That is a meaningful shift, not a threat. The scarce skill was never data entry. It was always the judgment to know what a number means for a business decision. Agentic AI removing the administrative layer around that judgment does not replace the analyst. It puts them closer to the conversations that actually mattered all along.

The common thread

Strip away the individual product pitches and the summit was describing one shift, over and over. TBM is moving from a discipline built on periodic reporting to one built on standing, governed, continuous systems, where agentic AI handles the repeatable work and humans stay firmly in charge of the judgment calls.

That is the same problem Yarken was built to solve: hardwiring AI into the semantic layer that already understands your technology spend, instead of bolting a chatbot on top of a report nobody trusts. If TBM Summit EMEA is any indication, that approach is quickly becoming the expectation, not the differentiator.