The Real Bottleneck Nobody Talks About

When a managing partner at a 50-person law firm deploys an AI tool, the expectation is speed. Faster document review. Faster client intake. Faster first-pass research. And the model itself, whether it is GPT-4 or Claude or something open-source, often delivers on that promise at the sentence level.

But workflows are not single sentences. They are chains of decisions.

Should this clause trigger a red flag? Does this financial summary meet the threshold for escalation? Is this client query routine or does it need a senior adviser? At every branch point, a traditional AI system pauses, calls a large model, waits for a response, processes the output, and moves to the next step. That cycle repeats dozens of times per workflow. The latency compounds. What looked like an efficient system on a demo starts to feel sluggish in production.

This is not a model quality problem. It is an architecture problem. And most firms buying AI tools have no visibility into it.

What Has Changed

A new class of decision-making model is emerging that treats routing and confidence assessment as a separate, specialised task rather than asking a large general model to do everything.

Jev is one example. It uses methods including parallel sampling and a training approach called RLCD to make discrete decisions extremely quickly, with the system reporting a speed improvement of up to 200 times compared to conventional chatbot-style decision-making. The model does not replace your large language model. It sits in front of it, handling the routing logic, and only escalating to a heavier model when confidence is genuinely low.

The practical consequence of that architecture is significant. For professional services firms running AI across multiple workflows, it means the expensive, slow parts of your AI stack get used less often, and only when they are actually needed.

What This Looks Like in Practice

Consider a senior associate at a mid-tier commercial law firm who is using an AI-assisted contract review system. The firm has 400-page agreements coming in regularly. The AI needs to assess each clause, decide whether it falls within standard parameters, flag anything that deviates, and assign a risk tier.

With a conventional setup, every one of those assessments routes through the same large model. The system is accurate, but it is not fast, and running 400 pages of clauses through it is expensive in both time and compute cost.

With a fast decision model handling the routing layer, the architecture changes. The vast majority of clauses, the boilerplate, the standard indemnities, the familiar structures, get assessed and triaged almost instantly. The heavy model is reserved for the genuinely ambiguous cases, the ones where a lower confidence score triggers escalation. The associate still gets the full review, but the system completes it in a fraction of the time and at materially lower cost.

That is not a theoretical outcome. It follows directly from how the architecture works. Fewer calls to expensive models. Faster triage. Same quality on the cases that matter.

The Implication for Firm Operations

Speed in AI workflows is not just a user experience issue. It affects how many matters a team can run concurrently, how quickly clients get responses, and how much your AI programme costs to operate at scale.

For a boutique accounting or financial consulting firm, the cost dimension matters most. Token usage, meaning the volume of text processed by large AI models, is a real line item in AI operating budgets. A routing layer that reduces unnecessary calls to expensive models is directly reducing that cost. You do not need to understand the underlying model to care about that. You need to read your AI vendor invoice.

For a law firm competing on responsiveness, the speed dimension is the one that creates competitive separation. Clients increasingly compare how quickly firms can turn around preliminary assessments, whether on a contract, a dispute matter, or a regulatory question. A firm whose AI infrastructure is genuinely faster at decision-routing will outpace competitors who are running the same models but in slower, less efficient configurations.

There is also a staffing question embedded here. If your AI system is slower than it should be, your people compensate by waiting or by doing manually what the system should handle. That is invisible overhead. It does not appear in your AI budget, but it appears in your utilisation reports and in how long your matter cycles take.

Our Take

Most professional services firms that have deployed AI are running on architectures that were designed for capability, not for operational efficiency. The focus was on what the AI could do. Less thought went into how decision-making inside the workflow was structured, and what that costs in time and money at scale.

Fast decision-routing models are a genuinely practical development. Not because they replace the models firms are already using, but because they change where in the workflow those models are needed. That changes the economics and the performance profile of the whole system. Firms that understand this architectural point will be able to build AI workflows that are faster and cheaper to run. Firms that do not will keep wondering why their AI investment has not produced the efficiency gains they were promised.

The capability gap between AI tools is narrowing. The operational gap between how firms deploy them is widening.

If you are running AI workflows in your firm and they are slower or more expensive than expected, the issue is almost certainly in the decision architecture, not the model.

Talk to us about how to audit and redesign your AI workflow for operational performance.