[ Thought Leaders
](https://www.unite.ai/series/thought-leaders/)
[Add Unite.AI to your preferred sources on Google](https://www.google.com/preferences/source?q=unite.ai)
Last year, Absa, one of the largest banks in Africa, needed 120 days to move a new financial crime detection model from build to production. While that was close to the industry average at the time, that same bank today can deploy a new model in 15 days, even with the same regulatory environment, transaction volumes, and compliance requirements.
And yet, a PwC survey found that 56% of enterprise CEOs have seen no significant financial benefit from their AI investments. Every enterprise leader I talk to has heard the pitch: if you deploy AI agents and automate your workflows, you’ll quickly see an ROI. Instead, 18 months and several million dollars later, the business impact is MIA. The missing ingredient is domain context, the operationally grounded understanding that gives an agent what it needs to make internal business decisions, and what to do when it errs.
When Agents Stop Advising and Start Acting #
Gartner estimates that by 2027, 50% of business decisions will be augmented or automated by AI agents for decision intelligence. Until very recently, AI’s role in the enterprise was to advise, leaving the employee to make the decision. The worst-case scenario was a missed insight, a delayed decision, or the lack of action in response to an alert. With agents, the consequences are much greater.
At a global pharmaceutical manufacturer’s biologics plant, when an inspector scans a lab door, the right inspection workflow launches automatically, tailored to that room’s equipment and specs. The agent sequences the inspection, surfaces the reference material, and lets the inspector confirm. In a GxP-regulated plant, every inspection must be audit-ready, meaning an agent is taking a share of the operational load in an environment with real consequences if it gets the sequence wrong.
Most governance frameworks in place today were built for the previous advisory world: periodic model reviews with a human checking a single system’s output. Now we have agents that hand tasks to other agents, coordinate across workflows, and compound a bad early call across an entire process chain before anyone notices.
Why Generic Agents Hit a Wall #
A general-purpose agent can look genuinely impressive under controlled conditions. It starts to struggle, however, in enterprise operations, which have constraints, interdependencies, and judgment calls that are obvious to anyone who has worked the floor and invisible to a model that has never been taught to see them.
A category manager at a top five global food and beverage company faced a version of the same problem on a different scale: 96% market penetration, with no traditional levers left to pull. The category manager defined the decision logic directly by unifying over 100 data sources across over 1,000 attributes per store and letting the AI model assemble stores into shopper behavior profiles that the category team could act on. The result? Three shopper clusters, weekly sprints with retail partners, an 8% incremental sales lift, and 35% faster shelf resets.
This worked because the agent already knew what a planogram is, what a shopper cluster means, and how a promotional change ripples into adjacent categories. A generic model must be taught industry-specific vocabulary, which costs 12 to 18 months most enterprises don’t have.
Why Governance Has to Be Architected #
The accountability frameworks most enterprises have today were built around explaining a model’s output after the fact and reviewing it periodically. They were not built for what happens when an agent triggers an irreversible action inside a workflow.
I’ve worked with an insurer operating across Central and Eastern European markets that has to consider each market’s regulatory requirements when it makes compliance decisions. Every customer’s due diligence decision is traceable as a live record that includes the input, model, output, and person signing off. The company’s AI system can route a transaction that’s compliant in one market and not in another without a compliance analyst’s manual adjudication.
This is the same architecture that lets Absa deploy its production AML system in 15 days instead of 120. The speed is built in, so the bank can immediately hand over the policy version for every decision to a regulator.
The enterprises we see investing in these governance layers early are still running agents at scale 18 months later. The ones that treat this as a compliance exercise bolted on after deployment are the ones having to keep rearchitecting.
Domain Intelligence Is Not Fine-Tuning #
Capable foundation models don’t reliably deliver enterprise outcomes on their own because they are built to be broadly useful. Enterprise operations require deep, narrow specialization. Conflating the two has already cost the industry time and resources, but it’s still the most common mistake I see.
Domain intelligence is not applying retrieval-augmented generation (RAG) or fine-tuning a general model. It’s the difference between an agent that can process information about a grocery store and one that understands how the business built and deployed 5,000 individually tailored planograms across over 750 stores without a central team touching every location.
This kind of operational fluency comes from years of domain-specific training data, decision logic built around real operational constraints, and a knowledge layer that includes how a pricing change affects margin, how an assortment shift affects shelf availability, and how that impacts shopper loyalty. For one large U.S. grocery chain, this connected reasoning produced $200M in incremental profit.
Building that grounding or getting it from a partner who’s already done the work is a real upfront cost in data, domain expertise, and the discipline to encode institutional judgment instead of skipping straight to a pilot. If that grounding is rushed, an agent that looks domain-aware in a proof of concept can still fall apart with the first production edge case.
What Getting Enterprise Agents Right Actually Means #
Enterprises that keep dropping capable models into existing workflows will have to keep admitting to their boards that the ROI isn’t showing up. MIT concluded last year that despite enterprise investments in generative AI, 95% of organizations are getting zero return. If you don’t want to be in that camp, put domain intelligence in place before day one, embed governance in the architecture, and hold agents accountable for specific outcomes.
The model behind any given agent is, ironically, the least durable part of that equation. A better one will ship next quarter, from the company that built the one you’re using today or one of its competitors. What doesn’t change with the next model release is whether your agents understand your business. You must build that in deliberately. The agents that succeed in your environment will not be the most capable, general-purpose agents on the market, but they will earn your trust to do more and more for your business.