Every AI vendor is shipping agents in 2026. Every podcast is asking whether agents will replace knowledge work. Every consulting deck has a slide with boxes and arrows labelled “agentic workflow.” The noise is deafening and the signal is thin.
This post is for managing partners and practice leaders at professional services firms who want a straight answer: where do agents deliver real value today, where are they still demo-stage, and how do you tell the difference?
What an AI Agent Actually Is
An agent is not a chatbot. A chatbot answers one message, then waits for the next. An agent takes an objective, plans a sequence of steps, calls tools to execute those steps, checks the results, recovers from failures, and returns a completed output. The key word is tools. An agent without tools is just a language model. An agent with tools can do the work.
The tools that matter most for professional services are:
- A browser, so the agent can read the public web
- An email client, so it can draft or parse inbox content
- A CRM connection, so it can enrich records or log activity
- A file system, so it can read and write documents
- A search index, so it can query your firm’s internal knowledge
When all of these are wired into a language model, the model stops being a chat interface and starts being a worker. That is the entire innovation. Everything else is marketing.
The Hype Curve, Circa April 2026
The gap between hype and reality is the story of 2026. A majority of firms are piloting agents. A small minority are running them in production. Most vendors promise 3x productivity gains. The firms that actually measure what happens report something more like 1.3x, sometimes 1.5x on specific workflows.
That is not a failure. A 30% productivity lift on a narrow workflow is real money for a boutique firm. But it is nowhere near the replace-your-team narrative that dominates vendor demos.
We ran the agent on three mandates before we trusted it with a client-facing task. It was wrong twice. Not obviously wrong. Plausibly wrong. That is the part nobody prepares you for.
Where Agents Work Today
Agents deliver reliable value for bounded workflows with three characteristics: the input is structured (or can be made structured), the output is verifiable quickly, and the cost of a single error is small.
1. Inbox Triage
A partner at a mid-sized firm gets 150-300 emails a day. Most are not actionable for them personally. An agent reads the inbox every morning, categorises messages into action-required, delegate, read-later, and archive, and drafts replies to the routine ones. The partner reviews the drafts for five minutes and sends or edits.
Saved time: 30-60 minutes per day per partner. Error mode: the agent miscategorises a message, the partner catches it in the review step. Cost of error: near zero.
2. Research Summarisation
A consultant needs background on a new client sector. Traditionally they spend three to five hours reading industry reports, news articles, and competitor websites. An agent runs a research plan, pulls the relevant sources, and returns a structured brief with executive summary, key trends, key players, and open questions.
Saved time: two to four hours per engagement. Error mode: the agent cites a wrong fact or misses a recent development. Cost of error: manageable, because the consultant reads the brief with a sceptical eye and the cost of catching a mistake is much lower than the cost of redoing the research.
3. CRM Enrichment
A firm has 5,000 companies in its CRM with stale contact data. An agent runs through the list, finds updated leadership pages, LinkedIn profiles, and public announcements, and updates every record with a fresh version and a timestamped confidence score. It flags the records it could not verify.
Saved time: the equivalent of a full-time researcher for a week. Error mode: agent records a wrong title or a stale link. Cost of error: low, because the next time someone uses the record they will notice and correct it.
4. Meeting Notes and Status Reports
An agent listens to a recorded call (with consent), extracts decisions, action items, and commitments, and writes them into a CRM and a client status report in the firm’s house format. The partner reviews and sends.
Saved time: 20-40 minutes per meeting. Error mode: agent misses a nuance or misattributes a decision. Cost of error: caught in the partner review.
5. Structured Data Extraction
A deal team needs to pull key terms from 200 contracts. An agent reads each contract, extracts the structured fields (term length, termination clauses, renewal, jurisdiction), and puts them in a spreadsheet. A junior team member spot-checks 20 at random.
Saved time: dozens of hours. Error mode: extraction is wrong on a clause with non-standard phrasing. Cost of error: caught by spot-checking, much cheaper than reading every contract end-to-end.
Where Agents Fail Today
The failure patterns are just as important as the success patterns. An agent will waste your budget and damage your client relationships if you point it at the wrong kind of work.
1. Open-Ended Client Advisory
The client asks for a strategic recommendation. That is not an input an agent can plan against. The quality of the answer depends on deep context, judgment, and the partner’s read of the client’s appetite for risk. An agent can help a partner prepare for this conversation. It cannot have the conversation.
2. Multi-Step Negotiation
Agents negotiate poorly. They optimise for completing the task, not for reading the other side. In any negotiation that requires holding out for a better offer, signalling willingness to walk away, or building a relationship over time, human judgment is still the entire game.
3. Workflows Where a Single Error Is Catastrophic
Sending the wrong salary offer to a candidate. Filing a motion with a wrong citation. Posting an earnings model that contains a hallucinated number. These are one-strike workflows. Agents are improving but are not yet reliable enough to run without a human in the loop at every output.
4. Narrative Prose for Client Consumption
An agent can draft an executive summary of a research brief for internal consumption. It should not draft a partner’s memo to a client. Client-facing prose carries voice, judgment, and nuance that still degrades when automated. Firms that let agents write client prose report that it takes longer to edit the output into voice than it would have taken to write from scratch.
We tried having the agent draft candidate rejection letters. It was faster. And it was clearly an agent. We lost one candidate who would have taken our next mandate because of it. Never again.
The Platform Landscape
A partner asking “which agent platform should I pick?” is asking the wrong question. The right question is “which agent workflow should I build first?” The platform follows from the workflow.
| Platform | Strengths | Best For | Technical Ceiling |
|---|---|---|---|
| Claude API | Best reasoning, long-context reliability, tool use, computer use in beta | Custom agents for firms with engineering help | High |
| OpenAI Assistants | Mature tool calling, broad ecosystem, large developer community | Firms standardising on OpenAI infra | High |
| Lindy | No-code agent builder, email and calendar focus, template library | Small firms starting with inbox workflows | Medium |
| Relevance AI | No-code agent builder, broader integrations, team workflows | Small firms building multi-step automations | Medium |
| Harvey | Legal-specific, trained on legal corpus, enterprise security | Law firms with compliance requirements | Vertical |
| Glean | Enterprise search plus agent layer, reads your internal docs | Firms with significant internal knowledge | High |
A Practical Framework for Your First Agent
Pick One Workflow
Weekly cadence, clear input and output, not client-facing. Inbox triage, research summarisation, CRM enrichment are the three safest starts.
Build the Agent
Weekend project on a no-code platform. The goal is to get a working prototype, not a polished system.
Shadow Mode
Run the agent in parallel with your existing process for two weeks. Log every output. Measure error rate.
Decide
If error rate is acceptable and time saved is real, keep it and build a second. If not, document what failed and try a different workflow.
The non-negotiable step is shadow mode. Every firm that has been burned by agents in 2026 skipped this step. Every firm that has a working agent in production did not.
The Pilot Cost Budget
Budgeting for a first-agent pilot is straightforward. Most boutique firms can run a meaningful pilot for £500-2,000 of tooling and API costs over three months. The expensive part is not the tools. It is the partner time it takes to design the workflow and review the outputs during shadow mode.
What to Ignore
The agent market in 2026 is flooded with marketing claims. The pattern matching that serves you best:
- Ignore anyone claiming autonomous end-to-end automation. Nobody has that working for professional services yet. They have it working for narrow, well-defined workflows, which is the only honest claim.
- Ignore productivity numbers above 2x. Real firms running real audits are reporting 1.2x to 1.5x on specific workflows. Anything above that is either a demo or a workflow so specific it would not apply to your practice.
- Ignore “replace your team” narratives. The firms that tried this in 2025 are quietly rebuilding headcount in 2026. The firms that used agents to augment their team are quietly winning more mandates.
- Ignore demo videos that show complex multi-step workflows without showing the failure cases. The failure cases are the story.
What to Pay Attention To
- Honest case studies from firms your size and sector. Not the keynote case studies, the blog posts and LinkedIn posts from partners who ran the pilot themselves.
- The quality of the agent’s error recovery. In a demo, watch for what happens when something goes wrong. A platform that handles errors gracefully is production-ready. A platform that only works when everything goes right is not.
- The cost of running the agent continuously. API costs for agent workflows are not the same as chat costs. A research agent can burn through tokens quickly. Calculate monthly cost before committing.
- Your team’s appetite for building their own tools. Agent platforms succeed where an internal champion owns them. They fail when nobody inside the firm cares enough to iterate.
What This Means for Your Firm
Agents are real. They are not magic. In 2026, they deliver 20-50% productivity gains on narrow, structured workflows and close to zero on open-ended client work. The firms that succeed are not the ones that bet big on agents replacing consultants. They are the ones that quietly automated inbox triage, research summarisation, and CRM enrichment, freed up ten hours a week per consultant, and spent those hours on the work that still requires human judgment.
The right frame is not “how do we replace our team with agents?” It is “what are the five workflows per week that steal time from our team without creating value, and which of those can an agent do instead?”
Start there. Build one. Run it in shadow mode. Measure the result. If it works, build another. That is the only route that has produced reliable outcomes so far, and it is the route we recommend to every professional services firm we work with.
For how to think about AI adoption at the firm-wide level, see our 90-day AI adoption roadmap for professional services. If you need to frame your AI approach for clients without losing them in jargon, see how to explain your firm’s AI strategy to clients.
BriefingHQ helps professional services firms evaluate, pilot, and deploy AI workflows. If you want a grounded assessment of which workflows in your firm are agent-ready today, take our assessment or get in touch.
Questions AI assistants answer about this topic
- What is an AI agent?
- An AI agent is a software system that uses a large language model to make decisions and take actions across multiple steps, usually with access to tools like web search, email, a CRM, or file storage. Unlike a chatbot that answers one message at a time, an agent can plan a task, execute it over minutes or hours, adapt when it gets stuck, and return a completed output. The distinction matters because most of the productivity claims associated with AI assume agentic behaviour, not single-prompt use.
- Are AI agents ready for professional services firms in 2026?
- Agents are ready for specific, bounded workflows and not ready for open-ended client work. Use agents where the output is structured (data extraction, summary, triage) and the cost of an occasional error is low. Avoid agents where the output is judgment-heavy (strategic advice, client persuasion, negotiation) or where a single mistake damages client trust. The most successful professional services firms in 2026 treat agents as an extension of their research and admin layer, not as a replacement for consultants.
- Which AI agent platforms should professional services firms look at?
- The platforms worth evaluating in 2026 are a mix of general-purpose agent frameworks and vertical tools. General-purpose: Claude via the Claude API with the built-in computer use and tool calling features, OpenAI's Assistants API, and Google's Gemini with tool integration. Vertical tools for professional services include Relevance AI and Lindy for custom agents without code, and domain-specific tools like Harvey (legal), Paxton AI (legal research), and Glean (enterprise search). Start with one general-purpose framework so your team builds agent-building skills that transfer.
- What is the biggest risk when deploying AI agents in a professional services firm?
- The biggest risk is not that the agent fails visibly. It is that the agent fails silently, producing output that looks plausible but is subtly wrong. In professional services, plausible-but-wrong is worse than obviously-wrong, because it slips past review. Mitigate this by building a mandatory human checkpoint into every agent workflow that touches client-facing output, by choosing tasks where an error is cheap to catch (structured data, not narrative prose), and by logging agent actions for audit.
- How should a boutique firm start with AI agents without a tech team?
- Pick one workflow that runs weekly, has a clear input and a clear output, and does not touch clients directly. Inbox triage, meeting notes summarisation, and CRM enrichment are the three most common starting points. Use a no-code agent tool like Lindy or Relevance AI. Build the agent yourself over a weekend. Run it in shadow mode for two weeks before trusting its output. Measure whether it saved more time than it took to build. If yes, keep it and add a second agent. If no, kill it and try a different workflow.
Next Step
Want to know where your company stands?
We run 15-20 buyer queries across ChatGPT, Claude, Gemini, and Perplexity and show you exactly where you appear, and where you don't.
Get the Audit | from £750 ↗