IIVPA/Knowledge base/Insight
Insight · AI consultancy

The real benefits of AI consultancy.

Almost every company now uses AI. Only a small share turn it into profit. We read the 2025–2026 research to find what separates the two — where an outside partner genuinely earns its fee, and where it doesn’t.

By IIVPA 6 November 2026 14 min read For CEOs, COOs, CFOs and transformation leads

The value gap, 2026

Using AI is near universal. Profiting from it is not.

80% say AI has made them personally more productive
37% report AI contributes anything at all to their organisation’s EBIT
6% are AI high performers, with 5% or more of EBIT coming from AI

Source: McKinsey, The state of AI in 2026, survey of 1,719 respondents, May–June 2026.

Five things to knowThe short version, if you read nothing else.

00The value gap

Using AI is easy. Profiting from it is not.

In 2026 the question is no longer whether a company uses AI. It is whether AI shows up in the numbers the board reads.

In McKinsey’s latest global survey, 80% of respondents say AI has improved their personal productivity. Yet only 37% report that AI contributes anything to their organisation’s EBIT — essentially unchanged from a year earlier — and just 6% qualify as “AI high performers”, attributing at least 5% of EBIT to it.1

The other large studies land in the same place. IBM’s survey of 2,000 CEOs found that only 25% of AI initiatives had delivered their expected ROI, and only 16% had scaled across the enterprise.6 S&P Global Market Intelligence found that 42% of companies abandoned most of their AI initiatives in 2025, up from 17% the year before.7

From evaluation to production

Of every 100 organisations, 60 looked at enterprise-grade AI systems. Five got one into production.

25%of AI initiatives delivered the ROI that was expected of themIBM CEO Study, 20256
46%of proofs of concept scrapped before production, at the average organisationS&P Global, 20257
>40%of agentic AI projects forecast to be cancelled by the end of 2027Gartner, 20258
Funnel: share of all organisations in MIT NANDA’s research at each stage, January–June 2025. Most stalled on brittle workflows, a lack of contextual learning and poor fit with day-to-day operations.5

MIT’s researchers put the cause plainly: the divide does not seem to be driven by model quality or regulation, but by approach.5 Which problem you pick, how you change the work around it, how you connect it, govern it and teach people to use it. That is the opening for an AI consultancy — and the test it has to pass.

The gap is closing — for some.

The companies pulling ahead share a handful of habits. Each benefit below is one of those habits, and what a good consultancy brings when an organisation can’t yet build it alone.

~50%of companies now generate measurable value from AI, up sharply on 20252
28%report revenue gains from AI, up from 20% in January 20264
01Benefit one

Start with the right problem.

The most valuable thing a consultancy does in week one is often to say no.

Budgets follow what is visible, not what pays. MIT found that around half of generative-AI budgets flow to sales and marketing, while some of the most dramatic, fastest-payback savings came from the back office: replacing outsourced processing, cutting external agency spend, automating risk checks.5

The market makes choosing harder. Gartner estimates that only about 130 of the thousands of vendors selling “agentic AI” offer genuine agentic capability — the rest is “agent washing” of chatbots and RPA — and warns that many use cases sold as agentic don’t need agents at all.8 BCG’s 2026 leaders, by contrast, start from strategic clarity about where AI creates value.2

Prioritising the backlog

Rank every candidate by value and by effort and risk — before anyone writes a line of code.

First workflowsWorth it, once the foundations existLow value for the effort
Illustrative example. Real rankings come from mapping your own processes, volumes and error costs.

What a consultancy brings

An outside team maps real processes with the people who run them, then ranks opportunities by value, risk and effort — independent of any one vendor’s product. You get a short, ranked backlog and an honest go / no-go, not a long list of pilots chasing whatever demoed best.

02Benefit two

Redesign the work, not just the tools.

This is the clearest difference between the companies that profit from AI and everyone else.

Nearly three-quarters of McKinsey’s high performers have fundamentally redesigned workflows because of AI, up from 55% a year earlier. Among all other respondents, it is one in four.1 Deloitte finds 60% of companies have redesigned jobs around AI to at least a moderate degree — but only 25% say they are using AI to deeply transform the business.4

BCG’s rule of thumb for AI leaders puts the technology in its place: 10% of the effort on algorithms, 20% on data and technology, 70% on people, processes and culture.3

Where the effort goes

The model is the smallest part of the job.

Algorithmsmodels and prompts
Data and technologypipelines, integration
People, processes and culturenew workflows, roles, training, incentives
Top: BCG’s 10-20-70 principle for AI leaders.3 Bottom: share that fundamentally redesigned workflows because of their AI use, McKinsey 2026 (each dot is 5%; “nearly three-quarters” shown as 75%).1

What a consultancy brings

Dropping a model into an unchanged process gets you a faster version of the old process. A consultancy redesigns the process around what the agent does well — who reviews what, where exceptions go, which steps disappear — and writes the new operating procedure with the team that will run it.

03Benefit three

Get to production, not pilot purgatory.

Experience is mostly a list of mistakes you no longer make.

In MIT NANDA’s interviews, AI delivered through external partnerships reached deployment about 67% of the time, against about 33% for tools built entirely in-house.5 Internal builds were far more common — and far less likely to ship.

Share of initiatives that reached deployment

Partnered AI shipped about twice as often.

MIT’s own caveat, and ours: the figures are self-reported, from interviews with 52 organisations, and may partly reflect the kind of company that chooses a partner. The gap was consistent across interviewees, but it is a correlation, not proof.
Source: MIT NANDA, The GenAI Divide: State of AI in Business 2025.5

What a consultancy brings

A team that has seen the failure modes: the pilot that only worked on clean demo data, the agent no one was allowed to connect to the ERP, the security review that started in week ten. It designs for production on day one — evaluation on your real historical cases, release gates, monitoring and a named owner — so the pilot is the first release, not a dead end.

04Benefit four

Make your data and systems ready.

An agent is only as useful as what it can read and what it is allowed to change.

In a Gartner survey, 63% of organisations either didn’t have, or weren’t sure they had, the right data-management practices for AI. Gartner predicts that through 2026, organisations will abandon 60% of AI projects unsupported by AI-ready data.9 And half of the CEOs in IBM’s study say rapid investment has left them with disconnected, piecemeal technology.6

What “ready” looks like

The unglamorous plumbing that decides whether an agent ships.

Data contracts
Scoped access per agent
Approvals and thresholds
Every action logged
Simplified architecture of a typical enterprise agent. The orange strip is what most failed pilots skipped.

What a consultancy brings

Before anything is built, it audits where the data actually lives, how clean it is and which systems an agent must read from and write to — including the old ones without an API. The fix is often less glamorous than the model: data contracts, access scopes, tested connectors and an audit trail.

05Benefit five

Build governance in from day one.

Autonomy is arriving faster than the controls that make it safe.

In BCG’s 2026 index, 42% of companies expect to give AI agents genuine decision-making authority by 2030 — but only 5% have the full set of controls in place today.2 Gartner predicts that by 2027, 40% of enterprises will demote or decommission autonomous agents because of governance gaps found only after production incidents. Its diagnosis: one rule for every agent fails both ways — over-restricting simple agents drives shadow development; under-restricting autonomous ones creates risk.10

Ungoverned use is already expensive. In IBM’s 2025 breach study, one in five organisations reported a breach involving “shadow AI”, and high levels of it added about US$670,000 to the average cost of a breach.12

42%vs5%expect agents with real authority by 2030, vs have full controls today2
40%of enterprises forecast to demote or switch off autonomous agents by 2027 over governance gaps10
US$670kadded to the average breach cost where shadow AI use was high12

Governance proportional to autonomy

Four levels of autonomy. Each earns its own controls.

Approve

The agent acts once a person approves — or alone, below a threshold you set.

The agent may
Post, pay or update records once approved; act alone under a value or risk threshold.
Controls
Thresholds and approval routes enforced outside the model; evaluation before every release; a record of each decision.
Who signs off
A named approver per threshold, with the agent’s evidence in front of them.
Example
Paying supplier invoices that match the PO and receipt, under an agreed amount.

In Hong Kong

The rules are mostly guidance built on existing law: the Privacy Commissioner’s Model Personal Data Protection Framework for AI (2024)15 and the Digital Policy Office’s Generative AI Technical and Application Guideline (2025),16 both resting on statutes such as the Personal Data (Privacy) Ordinance. Voluntary does not mean optional for a regulated business — it is what boards and auditors will measure you against.

What a consultancy brings

Governance set in proportion to autonomy — light controls for an agent that suggests, approvals and thresholds for one that acts — enforced outside the model and written to an audit trail your regulator can read. Built in from the first workflow, not bolted on after the first incident.

06Benefit six

Leave your people stronger.

Hong Kong’s workforce is already ahead of its training.

72.7% of professionals surveyed by HKUST use AI at work daily or weekly — yet nearly 60% have never received any AI training.13 IBM’s CEOs expect 31% of their workforce to need retraining or reskilling within three years.6

The upside is large when training lands. In a study of 5,179 customer-support agents, an AI assistant raised issues resolved per hour by 14% on average, and by 34% for novice and lower-skilled agents.14 In a randomised experiment with 758 BCG consultants, those using AI on tasks within its capabilities completed 12.2% more tasks, 25.1% faster, at more than 40% higher quality — but on a task outside that “jagged frontier” they were 19 percentage points less likely to get it right.17

Skills: the quiet bottleneck

Use is high. Training is not. And knowing where AI fails is a skill too.

+12.2%more tasks completed, inside the frontier
25.1%faster, inside the frontier
>40%higher quality, inside the frontier
−19 ptsless likely to be correct on a task outside it
Rings: HKUST Business School, 3,722 responses, March–April 2026.13 Bottom row: Dell’Acqua et al., field experiment with 758 BCG consultants.17

What a consultancy brings

Knowing where the frontier lies doesn’t arrive with a software licence. A good consultancy trains executives, operators and risk teams on your own platform and data, and hands over the runbooks — so the capability stays when the consultants leave.

07The fair warning

When consultancy makes things worse.

Not every outside partner closes the gap. Some widen it.

Gartner predicts that by 2028, 70% of enterprises will abandon agentic AI built by vendors’ forward-deployed engineers — bespoke systems that grow expensive to maintain and that no one inside can change once the vendor’s engineers leave. It also warns of “FDE washing”: consulting rebranded as engineering.11 And MIT noted that some of the clearest AI savings came from cutting spend on external agencies and consultants.5

The bar is rightly high. A consultancy is worth its fee when it leaves you owning the system and able to run it. It is a liability when it leaves you dependent.

Judge a consultancy by what you can do without it a year later.— The test we hold ourselves to

Leaves you dependent

  • Code, prompts and evaluations live in the vendor’s account
  • Success is measured in demos and pilot counts
  • One governance rule for every agent
  • Knowledge leaves with the team
  • Always recommends more AI

Leaves you capable

  • Deployed in your cloud; you own the code, prompts and evaluation data
  • Measured in hours returned, cycle time, accuracy and exceptions
  • Controls set by each agent’s level of autonomy
  • Runbooks, training and a named internal owner
  • Tells you where not to use AI
08How to choose

Seven questions to ask before you hire anyone.

Ask every firm on your shortlist — including us. Tick the ones they answered clearly, with evidence.

Shortlist check

Nothing is saved or sent. It only works out a score on this page.

0/7
Tick the questions a firm answered clearly, with evidence.
09A first engagement

What a good start looks like.

Start with one workflow and scale on evidence. This is how IIVPA runs a first engagement; timings vary with the process and the systems involved.

WEEK 0–2DiscoverMap processes with the people who run them; choose the first workflow.
WEEK 2–6BuildFirst agent built, evaluated on historical cases and secured.
WEEK 6–8LaunchLive on real volume, with people in the loop.
MONTH 3+RunMonitored, tuned and improved as processes change.
MONTH 6+ScaleMore agents, more desks — your team leading.
Hours returnedTime given back to people, per week
Cycle timeFrom request in to work done
AccuracyAgainst your own historical cases
Exception rateHow often a person has to step in

The four numbers we report from the first week.

10Questions

Common questions.

What does an AI consultancy actually do?
It helps an organisation decide where AI will create value, redesign the work around it, build and integrate it securely, govern it, and train people to run it. At IIVPA, consultancy starts with a two-to-four-week engagement that maps priority workflows and produces a ranked opportunity backlog, a target architecture, a risk assessment and a business case.
Isn’t it cheaper to build in-house?
Sometimes — if you already have experienced AI engineers, clean data and a mandate to change processes. In MIT NANDA’s sample, external partnerships reached deployment about 67% of the time against about 33% for internal builds, with the caveats above.5 Often the best model is hybrid: a partner builds the first system with your team, then hands it over.
How long before we see value?
A focused first workflow can be live on real volume in about two months: discovery in weeks 0–2, a first agent built and evaluated by week 6, launch by week 8. Enterprise-wide impact takes longer — and comes from redesigning workflows, not from running more pilots.1
We already use ChatGPT or Copilot. Do we still need this?
Keep them. General tools raise individual productivity, but MIT found they rarely move the P&L on their own.5 Consultancy focuses on the workflows where an agent connected to your systems changes cost, speed or quality — and on the governance that lets it act.
How does this work with Hong Kong regulation?
Hong Kong relies mainly on guidance built on existing law — the PCPD’s Model Personal Data Protection Framework for AI and the Digital Policy Office’s Generative AI Technical and Application Guideline, resting on statutes such as the Personal Data (Privacy) Ordinance.1516 Financial regulators such as the HKMA and SFC add their own expectations. A consultancy should map each agent against them and produce the evidence auditors ask for.

Bring one process. Leave with an honest answer.

A free 90-minute working session with an IIVPA partner and an engineer. You bring one process; we tell you what agents can — and can’t — do with it.

Book a working session How we work
11Sources

Sources and notes.

  1. 1McKinsey & Company, “The state of AI in 2026: On the road to ROI”, August 2026. Online survey of 1,719 participants, 4 May – 8 June 2026. mckinsey.com
  2. 2Boston Consulting Group, “AI Is Starting to Pay Off. Almost 50% of Companies Now Generate Value with It.” (Applied AI Index 2026), press release, 30 September 2026. 1,330 senior leaders. bcg.com
  3. 3Boston Consulting Group, “Are You Generating Value from AI? The Widening Gap”, September 2025 (10-20-70 principle; 5% of companies “future-built” for AI). bcg.com
  4. 4Deloitte, “Enterprise AI Enters an ROI Reckoning: More Use, Uneven Value”, press release on State of AI in the Enterprise, 29 October 2026. 3,500 leaders in 25 countries, May–June 2026. deloitte.com
  5. 5MIT NANDA, The GenAI Divide: State of AI in Business 2025, July 2025. Review of 300+ public AI initiatives, interviews with 52 organisations, survey of 153 senior leaders, January–June 2025. Preliminary findings; self-reported. report (PDF)
  6. 6IBM Institute for Business Value, 2025 CEO Study, 6 May 2025. 2,000 CEOs in 33 countries and 24 industries, February–April 2025. newsroom.ibm.com
  7. 7S&P Global Market Intelligence survey of 1,000+ respondents in North America and Europe, as reported by CIO Dive, 14 March 2025. ciodive.com
  8. 8Gartner, “Gartner Predicts Over 40% of Agentic AI Projects Will Be Canceled by End of 2027”, press release, 25 June 2025. gartner.com
  9. 9Gartner, “Lack of AI-Ready Data Puts AI Projects at Risk”, press release, 26 February 2025. gartner.com
  10. 10Gartner, “Gartner Says Applying Uniform Governance Across AI Agents Will Lead to Enterprise AI Agent Failure”, press release, 26 May 2026. gartner.com
  11. 11Gartner, “Gartner Predicts 70% of Enterprises Will Abandon Agentic AI Built by Vendor Forward-Deployed Engineering by 2028”, press release, 29 September 2026. gartner.com
  12. 12IBM and Ponemon Institute, Cost of a Data Breach Report 2025, July 2025. ibm.com
  13. 13HKUST Business School, “HKUST Workplace AI Study Reveals Hong Kong Leads the World in Adoption Rate with Over 70% of Employees Using AI at Work”, 10 August 2026. 3,722 valid responses, March–April 2026. bm.hkust.edu.hk
  14. 14Erik Brynjolfsson, Danielle Li and Lindsey Raymond, “Generative AI at Work”, Quarterly Journal of Economics 140(2), 2025. NBER Working Paper 31161. nber.org
  15. 15Office of the Privacy Commissioner for Personal Data, Hong Kong, Artificial Intelligence: Model Personal Data Protection Framework, June 2024. pcpd.org.hk (PDF)
  16. 16Digital Policy Office, HKSAR Government, “Digital Policy Office releases Hong Kong Generative Artificial Intelligence Technical and Application Guideline”, 15 April 2025. info.gov.hk
  17. 17Fabrizio Dell’Acqua, Edward McFowland III, Ethan Mollick, Hila Lifshitz-Assaf, Katherine Kellogg, Saran Rajendran, Lisa Krayer, François Candelon and Karim Lakhani, “Navigating the Jagged Technological Frontier: Field Experimental Evidence of the Effects of AI on Knowledge Worker Productivity and Quality”, Harvard Business School Working Paper 24-013, 2023. ssrn.com

Figures are quoted as each source published them; most come from self-reported surveys, and surveys differ in who they ask and how they define value, so they are not directly comparable. Where a study shows an association, we say so rather than claim cause. Sources checked on 6 November 2026.