AI agents for business are useful on the steps of a workflow where the system has to decide what happens next. An AI agent consultant’s real job is to find those steps before anything is built.
How does an AI agent consultant work in practice?
In practice, an AI agent consultant’s engagement starts from business objectives and feasibility, not from a preferred model, vendor or framework. It classifies the proposed system by family first, before any orchestration is built. It ends in a decision, and that decision can legitimately be to adopt a simpler solution, fix prerequisites first, defer, or not proceed at all.
That framing matters because an agent is an orchestration layer, not the language model inside it. Production agentic systems today are narrowly scoped, human-supervised systems built around LLMs, not LLMs acting autonomously. A consultant engaged mainly to write clever prompts is working on the wrong layer, and prompt-only builds have no answer the first time the agent must call a tool, recover from a failed step or stay inside a bounded action space.
Isn’t an AI agent just an LLM with a good prompt?
No. An agent is defined by the environment it operates in and the actions it can take there. Artificial Intelligence: A Modern Approach (1995) defines an agent as anything that perceives its environment through sensors and acts upon it through actuators. A survey of LLM-based agents by Xi et al. splits an agent into three parts, brain, perception and action. The brain is primarily a large language model, with perception and action as the other two components. The model handles reasoning, planning and decision-making. The surrounding orchestration layer supplies tool use, memory and control flow.
The 2024 survey of LLM-based intelligent agents lists context-length limits, slow knowledge updates and the inability to directly use calculators, SQL executors or code interpreters among the obstacles LLMs hit on practical tasks. Databricks describes the implementation requirements of agentic AI as looking “more like distributed systems + security + product controls than a typical ML integration”.
Treating the agent and the model as one thing causes scoping errors. Tool interfaces need design time too: Anthropic reports spending more time optimising tools than the overall prompt when building its SWE-bench agent. The 2026 requirements-engineering paper on delegated autonomy notes that leaving out “keep trying” limits is a common source of runaway behaviour. Reliability compounds across steps: at 95% accuracy per step, a 10-step task drops to about 60% and a 100-step task to 0.6%. A model that looks excellent in a chat window can fail as an agent for purely arithmetic reasons. For a deeper look at the architecture itself, see our explainer on what makes an LLM agent more than a language model.
Where agents sit among symbolic, traditional ML and generative systems
An agent is an architecture, not a model family, so it does not occupy a slot next to “symbolic” or “generative”. AWS Prescriptive Guidance draws the line this way. Traditional AI is tool-centric and functionally narrow, focused on prediction or classification. Traditional software agents add autonomy but are often rule-bound. Agentic AI combines autonomy, asynchrony and agency. The modern change is in the agent’s cognitive core: LLMs replace the static plan libraries and hand-coded rules that symbolic agents depended on. Sapkota et al. position generative AI as the precursor. In their taxonomy, AI agents are modular, task-specific systems driven by LLMs, while agentic AI adds multi-agent collaboration, dynamic task decomposition and persistent memory.
Hybrids of LLMs and classical planners exist. LLM+P, for example, hands long-horizon planning to a classical planner through PDDL and uses the LLM to translate the problem in and the plan back out. Our working taxonomy of symbolic, generative and traditional ML systems covers the families in depth, and the comparison of agentic AI vs generative AI with examples covers where the two terms diverge.
Which steps in a workflow actually need an agent?
Steps whose number or order cannot be fixed in advance, and where the system must act on other systems and recover from surprises, are the strong candidates for an agent. Anthropic separates workflows, where LLMs and tools are orchestrated through predefined code paths, from agents, where the LLM directs its own process and tool use. Microsoft’s Cloud Adoption Framework is blunt: if the steps are clear, repeatable and rule-bound, use regular code or non-generative models, which are faster, cheaper and more reliable for fixed workflows. AWS makes the same split. Highly standardised, simple tasks are better served by traditional automation or RPA, while tasks that need contextual understanding favour agentic approaches.
Not every inefficiency is an AI opportunity in the first place. One published audit methodology separates information gaps, process gaps, capability gaps, interpretation gaps and control gaps. Only the interpretation gap points to a potential role for AI-assisted analysis.
Step-by-step classification table
| What the step looks like | Better-fitting family |
|---|---|
| Fixed steps, strict rules, repeatable | Regular code, rules or a non-generative model |
| Single-pass classification, summarisation or translation | A direct model call, no agent logic |
| Answering questions from a fixed document set, no tools | Classic retrieval-augmented generation |
| Known stage sequence, interpretation needed inside stages | A fixed workflow with narrow, schema-bound LLM calls |
| Next step depends on what the system finds; must act on other systems | An agent, usually a single agent with tools first |
| Approval, privacy or ownership unclear | Governance work before any automation |
The progressive crystallisation production study names three execution types for such steps. Type 3 is agent-orchestrated investigation, Type 2 is a fixed structure where the LLM only interprets, classifies or summarises, and Type 1 is pre-coded logic at zero token cost. Where several agents seem necessary, Microsoft’s Azure guidance asks for justification first. Multi-agent designs add coordination overhead, latency and failure modes, justified when prompt complexity, tool overload or security rules defeat a single agent. We cover that threshold in when multi-agent architecture adds value.
When an agent is the wrong family, and what to build instead
Where the steps are clear, repeatable and rule-bound, Microsoft points away from an agent. For such fixed workflows Microsoft calls regular code or non-generative models faster, cheaper and more reliable, and the two studies below measured what a deterministic workflow with bounded model calls gains on specific tasks.
The “Blueprint First, Model Second” paper codifies the procedure as source code that a deterministic engine executes, calling the LLM only for bounded sub-tasks. On the TravelPlanner benchmark it reached a 35.56% final pass rate against 18.00% for ATLAS on the same backbone, with 96.0% fewer constraint violations (11 against 275) and 10.2 steps per task against 14.0. The authors attribute the gain to deterministic enforcement rather than a richer specification. Knowing a rule such as “each restaurant at most once” is different from being structurally unable to break it. The relative advantage also grew with task length: on 7-day trips the workflow scored 24.6% against 7.8%.
In the HS tariff classification study, fixing the control flow gave interpretability by construction, lower variance and failures that can be traced to a specific stage. Error analysis found 102 of 158 four-digit errors (65%) at a single ranking stage, where the gold heading was retrieved but not ranked first. An open-weight 27B-class model agreed with the frontier model on 84.2% of four-digit decisions inside the same workflow, so the architecture did not depend on the most capable model.
Cost follows the same direction. In a cloud network operations system handling tens of thousands of incidents a month, the progressive crystallisation study reports that deterministic workflows grew from zero to 45% of executions over eight months. Per-incident agent cost fell by more than 70% while incident volume doubled. The authors report reproducibility of roughly 50% for agent-orchestrated runs against 100% for deterministic ones, and argue that safety and cost improved together rather than trading off.
The Blueprint authors state the honest limit: determinism pays off in proportion to how well defined the procedure is. The tariff authors say explicitly that they do not claim planning autonomy is bad in general. The progressive crystallisation study treats agent exploration as a discovery mechanism rather than a permanent execution model. Build the workflow, keep a model inside the stages that need judgement, and leave a bounded agent for the genuinely novel cases.
How does a feasibility audit decide whether a problem justifies an autonomous agent?
A feasibility audit first asks whether the outcome is plausible at current capability levels, then whether an agent is the simplest architecture that reliably meets it. Asking an AI to be smarter than the people it serves is a very different proposition from asking it to take over a repeatable, menial task, and the second is far more likely to show measurable ROI. Our own per-use-case work on generative AI in marketing found the same mix inside a single budget line: tasks models can automate today, tasks that quietly need super-human capability, and questions that need investigation first.
On TheAgentCompany benchmark, which simulates a small software company, the best agent completed 30% of tasks autonomously. Simpler tasks were often solved, while long-horizon tasks remained out of reach.
Diagnostic checklist: does this task justify an agent?
The 2026 requirements-engineering paper treats an agent as justified when a task meets most of these six criteria:
- Open-ended structure: the number or order of steps cannot be reliably predefined.
- Unstructured context: success depends on reasoning over natural language or ambiguous inputs that make rules brittle.
- Cross-system actions: the task must inspect, update or orchestrate external systems, not just generate content.
- Evaluable progress: explicit success criteria or intermediate checks let someone catch a bad trajectory before harm occurs.
- Bounded risk: sandboxes, reversible actions, rate limits, approval gates or human takeover can contain harm, specified before deployment.
- Advantage over baselines: the team can state in advance a measurable threshold at which the agent’s added complexity pays off, and a way to verify it.
The paper records this reasoning in an Agency Justification Record, completed before the architecture is chosen. A task that passes still needs explicit stop conditions: maximum attempts, uncertainty thresholds, deadlines and the conditions that force the system to hand back control.
Next comes a pilot. Microsoft recommends piloting the hardest steps first and treating each go or no-go gate seriously: if a pilot misses the agreed results, pause and reassess rather than push on. The audit’s verdict is then one of five outcomes: proceed to a controlled pilot, adopt a simpler or existing solution, prepare prerequisites, defer and reassess, or do not proceed.
What an AI agent consultant delivers, and what stays outside the engagement
The useful output of an agent engagement is a decision with its reasons, not a prototype that happens to run. The audit classifies the proposed system by family first, so any agentic build is scoped against whether the problem is genuinely agentic before orchestration is written.
Published consulting guides describe the deliverables in more detail. One vendor guide lists discovery workshops, a use-case prioritisation matrix, a data and systems audit, a maturity assessment, a risk and governance review, a pilot plan with defined success metrics, a build-versus-buy recommendation and a phased roadmap. The same guide says the data audit alone often changes project scope, sometimes shrinking it and sometimes revealing that data cleanup has to happen first. Another methodology bundles the output into an investment assessment and roadmap: an executive decision summary, a current-state assessment, prioritised opportunities with rejected alternatives, an economic assessment, a risk and human-control assessment and an implementation sequence.
On time and price, the vendor guide puts standalone consulting phases at 1 to 4 weeks. It puts full readiness assessments at $10,000 to $60,000 depending on scope, seniority and region. Read that as one provider’s published range, not a market rate.
Several things sit outside such an engagement. Prompt-writing as the main product is one of them, for the reasons above. The production build is another: it is a separate phase, and ongoing advisory is billed apart from it. An engineering audit can set out the risks and human controls, but it cannot stand in for domain authority on compliance.
A hospitality audit, step by step
Hospitality shows the classification at work. One published hospitality discovery audit runs for two weeks across front-of-house, back-of-house and revenue operations. It maps workflows and systems such as Opera PMS, Cloudbeds and Toast, then scores the opportunities. Its candidates include an agent that parses incoming RFP PDFs and drafts proposal packages, and extraction agents that match distributor invoices against purchase orders and flag price discrepancies. Put through the table above, the invoice case reads as mostly a fixed workflow: extraction needs a model, while matching line items to purchase orders is rule-bound. RFP drafting has more genuinely open-ended reasoning in it. For the sector itself, see our AI engineering for travel and hospitality page.
Measuring ROI on AI agents for business: cost per accepted task
The relevant ROI comparison is cost per successfully completed and accepted outcome, alongside changes in quality, risk and capacity. One audit methodology defines it as follows:
Cost per accepted task = (model + software + human review + correction + ongoing operational cost) ÷ correctly completed tasks
The same source keeps the value of released staff capacity separate from cash savings unless a credible route to a smaller budget is documented. This matters because a narrow focus on accuracy has left agents needlessly complex and costly, as Kapoor et al. (2024) found in their analysis of agent benchmarks. AWS notes that agents typically carry higher upfront costs but lower per-transaction costs than human operations, so break-even volume matters.
Headline industry figures are worth knowing but should be read with care. The vendor guide cited above reports a Gartner projection that more than 40% of agentic AI projects will be cancelled by 2027, largely over unclear ROI and weak risk controls. It also reports a median payback of about 5.1 months on successful deployments, attributed to BCG and Forrester research from 2026. Neither figure describes your workflow. The pilot gate does.
Should we scope the agent ourselves or bring in a specialist?
Teams that already document their processes, hold clear ownership and run evaluation harnesses can often write the justification record and stop conditions themselves. Databricks warns that if current automation is fragile or undocumented and accountability is unclear, agentic AI will amplify the problem. A specialist earns their fee when the family classification is contested internally, or when the hardest step needs a pilot nobody on the team has built before.
Other routes exist. An internal team can follow published design discipline such as the 12-factor agent principles for reliable, deployable agents. How a consulting engagement runs from scoping to handover is covered in how a structured AI consulting engagement works.
FAQ
What is an example of an agentic solution?
Support-ticket triage is a good example. The system reads the request, checks logs, tries a fix, checks the result and escalates only when needed, so the workflow changes based on what it finds.
What does autonomy mean when managing an agent AI system?
In the agent literature, autonomy means the agent operates without direct human intervention and has some control over its own actions and internal state. In deployment it is graduated. AWS lists four approaches, from fully autonomous through human-in-the-loop and co-pilot to human-led with agent support, and every one of them involves humans in some capacity.
