The Actual Problem

It was not a staffing problem or a budget problem. It was a reliability problem.

Firms put automation in place for tasks like client onboarding, file reviews, compliance checklists, or matter summaries. The tool would handle the first few steps correctly. Then, somewhere around the middle of a multi-step process, it would drift. It would produce an output that looked right but was not. Or it would complete the task with a critical step quietly missing. No error message. No flag. Just a confident, wrong result sitting in a file waiting for someone to catch it.

In a 30-person accounting firm, that kind of failure does not just waste time. It creates liability. A compliance checklist that skips a field, a client summary that pulls from the wrong period, a tax position that reflects outdated guidance. The person reviewing it at the end may not catch it either, because the output looks complete. That is the nature of this failure mode. It does not look broken.

So firms did what sensible firms do. They added more oversight. They assigned someone to check the automation. Which is a reasonable short-term fix and a poor long-term strategy, because you have now paid for automation and paid for the person reviewing the automation, and you are slower than before because there is a new handoff in the process.

What Changed

The underlying issue was not that AI could not handle professional services work in principle. It was that the models available through most of 2023 and into 2024 struggled to maintain what you might call task fidelity across long, multi-step processes.

Give a model a short, bounded task and it would perform well. Ask it to reason through a 20-step workflow, hold context from step one while executing step 15, apply consistent rules across a large document, and return a structured output that matches what you specified at the start, and it would frequently lose the thread. Not always. Not predictably. Just often enough that you could not trust it without supervision.

That is a fundamental constraint, not a configuration issue. You cannot prompt-engineer your way around a model that does not retain sufficient context or apply consistent reasoning across a complex task. Firms that spent months trying to tune their way around this problem mostly concluded that the tool was not ready for the work.

What has changed is that the models themselves have improved materially. Newer models maintain coherence across longer tasks, apply instructions more consistently, and are significantly less likely to hallucinate with confidence, the pattern where a model invents a figure or a source and presents it as fact. That last point matters enormously in legal, accounting, and financial consulting work, where a confident wrong number is worse than no number at all.

What This Looks Like in Practice

Take a senior manager at a mid-tier financial consulting firm. Her team produces client-facing reports that draw on regulatory guidance, internal financial models, and prior engagement work. Each report has 12 to 15 discrete components. Previously, trying to automate that production process meant the output was reliable for the first half and needed to be almost entirely rewritten from step eight onwards.

With models that can now maintain task fidelity across the full sequence, the calculus changes. The automation handles the full draft. A consultant reviews and edits, rather than rebuilding. The time saving is real. The error rate drops. And critically, the review step becomes genuine quality control rather than a safety net for a broken process.

The same dynamic applies in legal work. A senior associate reviewing a 400-page commercial contract for non-standard clauses is not just doing a read-through. They are applying a consistent set of criteria across a large volume of material, flagging deviations, and producing a structured output that feeds into advice. That is exactly the kind of task where previous automation tools degraded in the middle. It is also exactly the kind of task where improved models now hold up.

What This Means for How Firms Are Structured

Here is the implication that most commentary on AI in professional services glosses over.

When automation was unreliable, the rational response was to keep humans close to every step. That preserved accuracy but it also meant the efficiency gains were marginal. Firms adopted AI tools but did not change their underlying workflows very much, because the tools could not be trusted to run without close supervision.

When automation becomes reliably accurate across complex tasks, the conversation changes. You are no longer asking "where can we use this tool to save a few hours." You are asking "which workflows can we redesign around the assumption that this runs correctly." That is a different question, and it leads to different answers about staffing models, billing structures, service capacity, and what your people spend their time doing.

Firms that work through that second question now will build processes that are genuinely more efficient. Firms that wait will implement the same tools later and wonder why they are not seeing the gains their competitors are.

Our View

The reliability problem was real and it caused a lot of firms to write off workflow automation too early or to over-invest in oversight layers that neutralised the benefit. The honest answer is that for complex, multi-step professional services work, many 2023-era tools were not ready for unsupervised deployment.

That has changed. The more useful question for most firms right now is not whether AI can handle the work. It is which workflows are worth redesigning first, and what a trustworthy automated process actually needs to look like for your specific context.

If you are running workflows that broke down in previous automation attempts, it is worth revisiting them with a fresh assessment before assuming the ceiling has not moved.

Book a working session with ROOVOLT and we will map the specific workflows in your firm that are now worth automating.