Who Do You Call for GenAI? What a Generative AI Consultant Actually Does

What a Generative AI Consultant Does

Five duties define the boundary in this operating model. Use-case scoping identifies which workflow is even a candidate. Model selection matches the workflow to a model and an architecture. Guardrail design sets the rules an output must clear before release. Evaluation design builds the test set the workflow is measured against. Production handover packages the evidence and hands the running system to the team that owns it, the same point the Production Path calls Productionise. Prompt-writing services, generic content production and virtual-assistant or chatbot-operator work sit outside this boundary; the excluded neighbour is the person who operates a tool, not the person who decides which workflows the tool is allowed to touch. Team titles and named authorship for this page publish once confirmed. Written by Kevin Dam, Founder, Synthesised AI Pte. Ltd.

Why Prompt Skills Are Not the Job

The consultant classifies each candidate workflow by how much error the business tolerates, attaches evaluation evidence to that workflow before any build starts, and hands the production team a scoped boundary it defends to a reviewer months later. Each of those three actions produces an artefact a person inspects, a classification record, an evaluation result and a handover pack, not a judgement call buried inside someone’s prompt file.

The output of the role is a decision, not a tool. A scoped boundary and the evidence behind it are what a generative AI consultant hands over; a folder of prompt templates is not the deliverable, and a team that receives only prompts has not received the scoping work at all. Vendors who reduce the engagement to prompt tuning sell technique, not the boundary decision the business actually needs signed off. In Synthesised.ai’s experience, model choice and prompt wording turn over as tooling improves, while the workflow-ownership decision, once evidenced, tends to hold for the life of the system.

Hallucination, content unfaithful to or unsupported by the source material feeding a model, is a systematic failure across generative systems, not an edge case a better prompt quietly removes, according to Ji et al.’s survey of hallucination in natural language generation, published in ACM Computing Surveys. NIST AI 600-1, the Generative Artificial Intelligence Profile of the AI Risk Management Framework released in July 2024, names confabulation, a model producing confidently stated but erroneous content, among the risks unique to or worsened by generative AI, and prescribes risk-management actions rather than a prompting fix. Both sources establish that hallucination and confabulation are systematic risks requiring systematic evaluation rather than a prompting fix; that the scope decision and its evaluation, rather than the exact instruction wording, most determine whether an output is safe to ship is Synthesised.ai’s interpretation, not a conclusion either source draws.

Why prompt technique is a commodity

Prompt technique is largely a commodity because published methods, few-shot examples, chain-of-thought prompts and retrieval prompting, spread quickly across teams and get absorbed into vendor tooling. The durable expertise sits in deciding which workflow a large language model gets to touch at all, in setting the error bar it must clear before release, and in producing the evidence that the bar was actually met.

How the scope boundary gets its evidence

The scope boundary gets its evidence from an eval-set, a fixed set of test cases with pass criteria the workflow must clear before release. The consultant attaches the eval-set to the workflow, re-runs it on model or prompt changes, and keeps the result on file, so the boundary stays measured rather than asserted from memory. The same eval-set discipline carries into a full generative AI consulting engagement once scoping starts, where the baseline is version-controlled and regression-tested as it is expanded, rather than frozen once and left to go stale.

Scoping What LLMs Own

The tiering follows the same logic NIST’s AI Risk Management Framework (AI RMF 1.0, 2023) sets out for risk generally: context and intended use frame the workflow first, then measurable criteria track its risk from there. The Never-LLM tier is Synthesised.ai’s conservative choice to exclude the LLM entirely from certain decisions, not a regulatory requirement; Singapore’s Model AI Governance Framework for Agentic AI, which IMDA describes as a guide for enterprises and a living document (launched 22 January 2026, updated on 20 May 2026 with real-world case studies and new best practices), recommends risk-based boundaries and human-approval checkpoints for irreversible or high-stakes actions rather than a blanket prohibition, and where full exclusion is legally required it follows from sector-specific law. This four-tier classification is Synthesised.ai’s own framing.

Skills and Deliverables

Three of the four skills surface as distinct perspectives below: the technical work of choosing a model and its retrieval architecture, the evaluative work of proving accuracy rather than assuming it, and the organisational work of setting guardrails, sign-off and handover. The fourth skill, translating the same boundary so procurement, engineering and a regulator all read identical evidence, sits inside the organisational deliverable rather than standing as a separate document, because a handover pack that only engineering reads has not actually finished the translation work the role exists to do. Each skill produces an artefact a procurement reviewer opens without a translator, the practical test that separates a competent hire from a demo-capable one.

Technical: model and retrieval architecture

The technical skill selects the model and the retrieval architecture, then records the choice. A model choice made without a written rationale does not survive a staff change or a vendor price rise, so the record matters as much as the choice itself. Retrieval-augmented generation lets a model draw on an inspectable external knowledge source and can improve factuality over a closed-book baseline, per Lewis et al.’s Retrieval-Augmented Generation research presented at NeurIPS 2020; it does not by itself guarantee faithful answers or claim-level traceability, which need separate citation-generation and faithfulness-evaluation controls. Grounding matters most for workflows already classified LLM-drafted or LLM-assisted, where a person checks the output against a source before it ships. The deliverable is an architecture decision record and a model-selection rationale, two documents a buyer keeps and a future team audits without re-running the original evaluation.

Evaluative: test design and failure analysis

The evaluative skill designs the test set and reads the failures. Standardised safety evaluations for large language models cover robustness, factuality and bias, categories set out in IMDA/AI Verify’s Cataloguing LLM Evaluations work and Project Moonshot, and the consultant turns those categories into pass criteria on a version-controlled eval-set; Singapore’s Model AI Governance Framework for Generative AI (2024) separately calls for testing, assurance and common standards at the ecosystem level. NIST AI 600-1’s confabulation risk, covered above, gives the failure category the eval-set is built to catch, and every workflow classified above Never-LLM carries at least one pass criterion tied directly to that risk. A failed test case gets logged, not silently reworded until it passes, because a rewritten test proves nothing about the model that runs in production. The deliverable is an evaluation set with pass criteria bound to each workflow, so accuracy is a documented result rather than a claim made in a sales conversation.

Organisational: guardrails, sign-off and handover

The organisational skill sets the guardrails, the approval chain and the handover. It names who signs off each ownership tier and packages the evidence procurement and a regulator both read, so a review months after go-live starts from a record rather than from memory. The deliverable is a guardrail policy, an approval chain and a production handover pack, and the same pack carries the translation work, written so procurement, engineering and a regulator read identical artefacts rather than three separate summaries. A scoped workflow that clears its evaluation bar hands over to an AI agent build once the ownership decision calls for autonomous, multi-step execution rather than a single-pass output. Client-specific deliverable names and templates publish once confirmed.

When to Engage One

A generative AI pilot running with no evaluation set is the first trigger, because a pilot without a frozen test set has no way to prove it worked before it reaches more users. A workflow shortlist with no safety boundary assigned is the second, a list of candidate use cases sitting unscoped while budget pressure pushes one of them into production anyway. A tender or an audit asking for governance evidence the team cannot produce is the third, the moment a procurement reviewer or a regulator asks for a document that does not exist yet.

MIT NANDA’s State of AI in Business 2025 report (July 2025) is a preliminary, interview-based report whose production figures it describes as directional; it estimates that about 95% of organisations obtain no measurable profit-and-loss return from generative AI, that around 5% of integrated pilots extract substantial value, and that around 5% of task-specific systems reach production. That a missing eval-set is the cause is this page’s inference, not a finding of the report; it does support the first trigger, since a pilot without a frozen test set has no way to prove it worked. The same workflows a generative AI consultant scopes usually map through an AI workflow automation build once the ownership decision clears them for automation rather than a single generative step.

From Role Definition to a Scoped Engagement

Because the ownership decision precedes any build, a generative AI consulting engagement scopes the workflow map, the guardrails and the evaluation evidence in one piece of work. This page defines the role; the generative AI consulting service scopes the workflow map, the guardrails and the evaluation evidence in one engagement once a business is ready to act on it. Every commercial term after scoping arrives in a written proposal, never as a price on this page. The proposal names the engagement model that fits the scope, Discovery & Roadmap, Fixed-Scope Build, Managed Build-and-Run, Advisory Retainer or Fractional AI Leadership, and every route carries zero published figures. Enquiries route through the enquiry form on that service page.