AI Agents in Professional Services: How consulting firms are changing the way they deliver

Professional services delivery is under pressure. The 2025 SPI Benchmark put PS firm EBITDA at 9.8% — the lowest in five years. Billable utilization fell to 68.9%. On-time delivery sits at 73.4%. The work has not gotten easier. The margins have gotten worse. And the headcount math no longer works: you cannot hire your way out of utilization gaps when every new consultant takes six months to ramp.

This is the gap AI agents are filling. Not as chatbots. Not as copilots suggesting edits. As autonomous task executors — assigned to project work, operating within defined guardrails, producing structured deliverables that humans review before anything moves downstream.

In 2026, AI agents in professional services are autonomous task executors that handle the repetitive operational work — scoping, status reporting, time capture, meeting recaps, onboarding execution — while humans keep ownership of strategy, client relationships, and approval authority.

This article covers what is actually happening in PS delivery, the categories of work agents handle today, and how to evaluate whether your operations are ready for this shift.

What Does an AI Agent Actually Do in a Services Context?

An AI agent in professional services is software that executes a specific project task autonomously within defined guardrails. It is not a chatbot that waits for a prompt. It is not a copilot that suggests edits. An agent runs on a trigger, produces a structured output, and routes that output to a human for approval.

The key distinction: agents act, then humans approve. The human does not prompt the agent. The agent fires on a project event — a deal closing, a week ending, a milestone approaching — and delivers output to an approval queue. The human reads, edits if needed, and approves or rejects before anything moves downstream.

The best implementations treat agents as project resources, not as a separate AI layer. The agent is assigned to a task. It appears in the same plan as the human team. The PM supervises it the same way they supervise a consultant — by reviewing what it produces and deciding what ships.

What Categories of PS Work Do AI Agents Handle Today?

The agents in production today cluster into a handful of repeatable PS workflows. These are the categories where the work is structured enough for an agent to execute, and where the volume of work justifies the investment.

For a concrete view of agents running today, see the Klient PSA agent catalog.

What Is the Hybrid Project Delivery Model?

Hybrid project delivery is the operating model emerging in firms that have moved past AI experiments and into production. It is a permanent structure where human consultants and AI agents share a project team — with explicit roles, structured handoffs, and human approval gates at every output.

The division is clean. Agents handle repetitive execution: drafting scope, capturing time, compiling status reports, monitoring milestones, flagging risks. Humans lead with strategy, client relationships, architectural decisions, and approval authority. Every agent output passes through a structured approval gate where a human reads, edits, and approves before anything moves downstream.

The model only works if the agent is treated as a project resource, not as a parallel system. The agent appears in the same plan, the same resource view, and the same approval workflow as the human team. The PM does not learn a new tool to manage agents — the agents show up where the work already lives.

Klient PSA is one example of this model in production. Agents are assigned as Salesforce Resource records, appear in the same Gantt charts as consultants, and route output through the Outcome Review object for human approval. The pattern is what matters — the implementation will look different on different platforms, but the principles do not change.

Security has to be built into the architecture, not bolted on. The strongest implementations run agents through an enterprise trust layer that enforces data masking, toxicity filtering, prompt injection defense, and zero data retention by LLM providers. Salesforce Agentforce provides this through the Agentforce Trust Layer: project data never leaves the Salesforce security perimeter, and agents inherit the same profiles, permission sets, and sharing rules as the users they serve. There is no separate AI security model to configure or audit — it is the same security the org already runs.

The 2025 SPI Benchmark found that firms using Gen AI report 5.4% higher revenue growth than firms that do not. Hybrid delivery is the operating model that turns that stat from a survey answer into a daily practice.

What Does This Mean for Consulting Firm Operators?

Hybrid delivery changes the rhythm of project work, not the org chart. Here is what shifts and what stays the same.

The net effect: your people spend more time on the work that requires judgment and less time on the work that requires assembly. A PM who spends 3.7 hours per week rebuilding project context gets that time back. A consultant who loses 20% of their billable hours to manual timesheet entry gets those hours recorded. The margin improvement comes from recaptured time, not from reduced headcount.

How Do You Evaluate If Your PSA Is AI-Ready?

Not every PSA can run AI agents. The platform architecture determines whether agents are a native capability or a bolt-on experiment. These five criteria separate AI-ready platforms from the rest.

1. Is it on a platform with native AI infrastructure? Salesforce Agentforce is the native AI agent platform for Salesforce. A Salesforce-native PSA runs Agentforce agents directly on project data — no middleware, no API bridge, no separate AI service. If your PSA is standalone SaaS, agents require custom integration.
2. Can agents access project data without middleware? Agents need to read and write project records — tasks, timesheets, resource assignments, billing rules. A native PSA gives agents direct access to the same data model the PM uses. An integrated PSA requires the agent to call an API, wait for a response, and handle sync conflicts.
3. Does the PSA have structured approval workflows? AI agents should never ship output without human review. The PSA must include an approval gate — like the Outcome Review in Klient PSA — where humans see agent output, edit it, and approve or reject it before anything reaches a client or a downstream system.
4. Can agents be assigned as resources alongside humans? In Klient PSA, agents are assigned to project tasks as Salesforce Resource records. They appear in the same Gantt charts and resource views as consultants. This is not a cosmetic feature — it means PMs manage agents with the tools they already know.
5. Does the AI layer enforce enterprise-grade security and trust? Ask where the data goes when an agent runs. The Salesforce Agentforce Trust Layer enforces data masking, toxicity filtering, prompt injection defense, and zero data retention by the LLM provider. Agents inherit Salesforce profiles and sharing rules. If the vendor’s AI runs on a separate infrastructure with a separate security model, your compliance team will have questions that do not have good answers.
6. Is the vendor running agents on their own operations? Customer Zero credibility matters. Klient ran every agent on its own delivery projects before releasing them to customers. If the vendor is selling agents they have not used themselves, the maturity gap will land on your team.

Klient PSA checks all six. It runs on Salesforce, uses Agentforce natively with the Trust Layer, includes the Outcome Review for every agent output, assigns agents as resources, and operated as Customer Zero for every agent in the PSA catalog.