AI can now find information, propose a decision and, increasingly, act on it. Whether that earns money for your business depends on three things the model cannot supply: evidence you can explain, experience you have kept, and clear limits on what the system may do.
Andy Demir | Qalbur Perspective | Future of Work | 15 minute read
Executive Summary
AI can make individuals more productive without making the business perform proportionally better. Putting Bernard Marr’s business-led approach to data strategy alongside Schank, Lyras and Soloway’s case for relevant experience shows what is missing: your decisions need dependable evidence, the context to interpret it and clear, accountable authority to act. Current research supports preparing around specific use cases rather than waiting for perfect enterprise data. My suggestion is a simple decision contract that ties these together, so you can test bounded automation while keeping human judgement, organisational memory and an honest account of where the commercial value comes from.
In This Article
- When records become decisions
- Two books, two different disciplines
- Productivity does not automatically become performance
- What practitioners are questioning
- Different starting points across Europe and the Gulf
- What the company remembers, and what it leaves out
- A decision contract for bounded automation
- The case against waiting for perfect data
- What Senior Leaders Should Consider
- Questions for the Boardroom
- Reasoning worth preserving
- References
When records become decisions
Take a purchase order, a sales forecast and a customer complaint. Each can describe the same business in a different way, and each may be accurate for its own purpose. The trouble starts when an AI system combines them and hands you an answer without knowing which definition, which time period, or which commercial obligation governs the decision before you. A polished recommendation can hide a disagreement your organisation has never actually settled.
That is why the old line, before AI strategy comes data strategy, still earns its place in 2026, as long as you treat it as a discipline you practise decision by decision, rather than a programme you must finish first. AI readiness means evidence fit for the decision in hand, enough context to interpret it, and clear authority over what happens next. Putting every database in one place is neither the whole answer nor always the first step. Start with the business decision you want to improve.
Agents raise the bar again. By an agent I mean a system that can work through a task in several steps, use tools and, within its permissions, change something outside the conversation. An assistant that drafts a supplier comparison leaves you to decide what happens next. An agent allowed to place the order also has to recognise contractual conditions, spending limits and the moments when it should stop and ask. Two older books, written well before anyone was talking about agents, help explain what that demands of an organisation.
Two books, two different disciplines
It may seem odd to reach for two books written before generative AI arrived, but that is part of their value. They describe the management problem without the claims surrounding today's tools, and they approach it from opposite ends. Bernard Marr writes for business leaders, asking what data a company should collect and for what purpose. Roger C. Schank and his co-authors come from cognitive science and software, and ask what experience a person needs at the moment of decision. An agent acting on your records needs good answers to both questions, which is why I have put the two books side by side here. Neither offers a manual for AI, but together they show you what a system has to know before you let it act.
Marr's Data Strategy: How to Profit from a World of Big Data, Analytics and the Internet of Things (Kogan Page, 2017) offers the first discipline. A business and technology writer and adviser, he ties data collection to business objectives, and his most durable point is a plain one: decide what you need to know before you decide what to collect. His examples and forecasts from that time cannot tell you much about what today's AI can do. [1]
The Future of Decision Making: How Revolutionary Software Can Improve the Ability to Decide (Palgrave Macmillan, 2010) supplies the second. Schank, a cognitive scientist, wrote it with Dimitris Lyras, a maritime software entrepreneur, and Elliot Soloway, a researcher in learning technologies. They argue for software that brings relevant experience to a decision. Their approach, case-based reasoning, looks for useful similarities in earlier situations and then tests whether those lessons still apply. Goals and conflicts are what tell you which experience is relevant. [2]
Read side by side, even in summary, the two books expose a gap in any strategy built on data alone. A record of what happened often says little about why a competent person chose it, which alternatives were rejected or what remained uncertain at the time. Marr helps you establish purpose; Schank and his co-authors help you recover context. My argument is that an AI-ready decision needs both, and it needs to be connected to someone who is accountable for the action.
Neither book is a technical blueprint for generative AI. You should also be wary of the decision-making book’s suggestion that software removes emotional bias, because what gets selected, recorded and retrieved can carry bias just as easily. Marr’s book has moved on as well: a third edition appeared in September 2025 with a new subtitle. What follows starts from summaries of the earlier editions and does not pretend they describe the latest ones, or that every claim in them has survived unchanged. [3]
Productivity does not automatically become performance
The recent evidence supports preparation, though you need to handle the headline adoption figures with care. Fivetran’s 2026 Agentic AI Readiness Index, published in May 2026, is based on a survey of 400 data professionals at mid-sized and large enterprises in the US, UK, wider EMEA and Asia-Pacific. It found that 41% of organisations were already using agentic AI in production while only 15% were fully prepared for it. This is vendor-sponsored research in a defined sample of larger organisations. Still, it puts a number on a tension you may recognise: agents are going live faster than the data foundations they depend on. [4]
McKinsey’s State of AI survey, published in August 2026 and based on 1,719 respondents across 97 countries, shows a different gap. Eighty per cent said AI had improved their individual productivity, while 37% attributed at least some EBIT impact to AI, a share McKinsey describes as essentially unchanged from the year before. These are different self-reported measures, and the gap between them does not explain itself. It does warn you against treating faster individual work as proof that the business has captured the value. [5]
Experimental research lends more weight to that warning. In a six-month randomised field experiment across industries, half of 6,000 workers received a generative AI tool built into the applications they already used. The tool changed what people could adjust on their own, such as time spent on email, far more than work that needed coordination, and time in meetings did not change significantly. The study is a 2025 working paper and predates the latest agent deployments, so it cannot settle what agents will do. It does suggest why better tools can leave approval queues and cross-functional decisions largely where they were. [6]
More complete data will not remove every such delay. A margin decision can rest on accurate information and still involve a legitimate disagreement between winning an important customer and protecting capacity for more profitable work. Settling that takes executive judgement about priorities. Hide it inside an instruction to maximise sales, and an agent will execute efficiently while moving your business away from what you actually intended.
What practitioners are questioning
Practitioners’ public discussions bring these tensions closer to daily work. An April 2026 Reddit thread about agent-generated dashboards raised worries about ambiguous definitions, permissions and small errors hidden behind convincing presentation. A January thread about data-engineering careers asked what is still valuable once AI can complete a neatly specified tutorial project. Threads like these show you which questions are worth investigating; they cannot tell you how common the problems are or what will happen to jobs. [7] [8]
What separates a useful answer from a dangerous one is whether anyone knows it answers the business question. An analyst can generate a perfectly valid query for active customers and still leave out customers whose contracts are live but whose buying is seasonal. The arithmetic is right and the commercial reading is wrong. So your data strategy needs shared definitions with an explicit scope, including the places where different definitions are legitimately required.
Professional commentary on LinkedIn puts growing weight on context, data lineage and validation for agents. Data lineage means being able to trace a figure back through its sources and the transformations it went through. Some of that commentary comes from people selling the tools that promise to solve the problem, which limits its value as evidence. A commercial interest does not make the question wrong, but you should verify product claims separately. [9]
All of this points to a more useful management response than asking employees to write better prompts. Finance and commercial teams can agree which revenue definition governs which decision, and data specialists can make that agreement usable inside the system. Your people also need a recognised way to challenge an AI answer and get it corrected. Without one, informal checking quietly becomes permanent work, and its cost never appears in the business case.
Different starting points across Europe and the Gulf
How much this work pays back depends on where you start. Eurostat’s Digitalisation in Europe 2026 edition reports that 20% of EU enterprises employing at least 10 people used AI technologies in 2025. Use ranged from 42% in Denmark to 5% in Romania, with Belgium at 35% and Poland at 8%. These figures measure whether AI is used at all. They say nothing about data quality, agent autonomy or executive capability. [10]
Even so, they should make you wary of a single European template. A company adding agents to established digital workflows faces different choices from one that still reconciles operational records by hand. National averages tell you nothing certain about an individual firm, and low aggregate adoption does not rule out advanced local deployments. For a smaller business, a single, dependable knowledge base for customer service may be a better investment than an enterprise architecture designed for a much larger company’s needs.
A Spanish banking group gives a concrete example from Italy and Germany, where BBVA runs its two fully digital banks. In July 2026 BBVA reported that generative AI assistants used by its contact-centre agents had cut average handling time for common enquiries about products, services and procedures by more than 15%. The tools draw on product and procedure knowledge, have no access to customers’ personal or sensitive data, and produce draft responses that agents review before sending. Questions that depend on an individual customer’s account stay with the human agent. The reported benefit comes from a tightly bounded task. [11]
Poland offers a different test of where the boundary sits. In July 2026, PKO Bank Polski reported the first purchases initiated by an AI agent on behalf of its cardholders. The purchases were made with Visa at real online shops participating in Visa’s European Agentic Ready programme. The bank stressed cardholder consent and control, as well as the use of existing payment safeguards. This was a programme test, and you should not read it as evidence of widespread autonomous commerce. What it shows is that the authority to initiate a purchase can be defined separately from the intelligence used to find it. [12]
A Gulf example raises questions about language and interpretation. Emirates NBD describes a platform for analysing customer conversations to surface recurring service problems. Its account states that the platform supports English, with Arabic and other regional languages planned. This is the bank’s own description, and nobody has measured the outcome independently. It is a useful reminder to test a system’s apparent coverage against the languages and circumstances of the customers whose experience it claims to represent. [13]
Context, in other words, is part of the evidence. A complaint classified as a request for information may reflect an earlier interaction nobody resolved, and a supplier recorded as late may have been working to a revised delivery date. Both examples are hypothetical. They still show why retrieving more documents helps only when the system can tell which version and which explanation belong to the decision.
What the company remembers, and what it leaves out
This is where the second book earns its place. A well-structured case keeps the objective, the alternatives on the table, and the conditions under which a decision was made. A supplier substitution that protected production during one disruption may have depended on a temporary approval or an unusually tolerant customer. Reuse the outcome without those conditions and useful organisational memory turns into misleading precedent.
Experience needs scrutiny too. A collection made up mostly of successful decisions will flatter established practice, and a confident explanation written afterwards can hide what nobody knew at the time. Your best defence against hindsight is to record what was known before the decision and review it separately from the outcome. A commercially sound decision can still produce a poor result under uncertainty, and a lucky result can hide weak reasoning.
Capturing this material depends on trust. If every lesson is immediately turned into a judgement about the people involved, your employees have every reason to offer tidy accounts with no uncertainty or disagreement. If you want useful memory, explain how cases will be used, protect sensitive material and recognise the people who contribute to learning. At that point your AI project is tied to a harder question: can your organisation hear an honest account of its own work?
There are consequences for capability as well. A study presented at CHI 2025, based on a survey of 319 knowledge workers, found that greater confidence in generative AI was associated with less reported critical thinking. It is a survey and does not prove that AI causes a lasting loss of ability. It does pose a design question: will your employees have the knowledge, time, and authority to examine a recommendation, or will review become an approval ritual? [14]
A decision contract for bounded automation
A practical way to bring these concerns together is what I would call a decision contract. I mean a short, maintained agreement about the decision an AI system supports, the evidence it may use, the context it needs and the actions it may take. It is a management tool with no legal force, and it is no new technical standard either. Its value is that it makes unresolved assumptions visible before they turn into automated behaviour.
Take supplier selection. The contract would say whether the objective is the lowest purchase price, dependable delivery or total cost over the life of the contract. It would define how quotations are compared and who confirms that delivery information is current. Relevant cases would record the exceptions behind earlier choices, and a named business owner would settle the conflicts the instructions cannot resolve.
Authority belongs in the same agreement. You might allow the system to prepare a comparison, request missing information, or recommend a supplier, while any financial commitment requires a defined approval. A stale quotation or a contradictory specification should trigger a request for clarification, never a completed transaction. Where action is permitted, you need a record of what changed and a workable way to correct it.
Testing then asks whether the arrangement survives ordinary complications. The test cases should include duplicate supplier identities, conflicting versions of a document and requests outside the intended scope, alongside routine tasks that should succeed. Anthropic’s engineering guidance on evaluating agents, published in January 2026, recommends combining automated evaluations with production monitoring and periodic human review. Guidance is no proof that any particular system is safe. It does support judging an agent on its actual performance throughout deployment, long after the impressive demonstration is over. [15]
The case against waiting for perfect data
The strongest objection to all of this is that it could become just another expensive governance exercise. You cannot wait until every record is clean and every disagreement documented. AI itself can help find inconsistencies, draft definitions and recover relevant material from neglected files. For low-risk tasks, controlled experimentation is often the fastest way to find out which preparation actually matters.
Salesforce’s own research supports that objection in part. Its survey of 2,025 agentic AI decision-makers in 20 countries was fielded in May 2026 and published that August. It found that only 31% of the organisations that had deployed agents had fully unified their data before launch. Among those deployers, organisations that unified the relevant data beforehand reported achieving a meaningful return on investment sooner: in 7.3 months, compared with 8.8 months for those that fixed infrastructure gaps afterwards. These are self-reported associations published by a vendor, and they do not prove causation. They support a narrower conclusion: preparing the data a particular use needs can pay off without complete enterprise integration. [16]
So ‘before AI strategy comes data strategy’ describes an order of reasoning, and reading it as a rigid project sequence loses most of its value. Business purpose tells you what data you need, and experiments tell you whether the proposed use is worth it. Data preparation and AI development can move in tandem around a single bounded decision. The real trade-off is how much uncertainty you accept in exchange for learning sooner. My view is that the limits should tighten wherever an error creates obligations that are hard to reverse.
What Senior Leaders Should Consider
You can make that trade-off concrete by choosing one recurring decision with a visible business consequence. Repeated customer contacts, slow quotations or avoidable stock shortages are better starting points than a general ambition to become AI-led. The business owner can establish the current process and its cost, while the people who do the work identify the exceptions the official procedure omits. Measure the first deployment against that baseline.
Your financial case must include preparation, supervision, and correction. Faster quotation preparation is of little value if approval remains the bottleneck, and the minutes saved in service need to be weighed against repeat contacts and resolution quality. Released capacity may improve resilience or let you serve more customers without reducing spending at all. If you want to call it a saving, explain how the benefit actually reaches the business.
Treat human capability with the same discipline. Experienced employees can help curate cases and interpret unfamiliar results, while less experienced colleagues practise forming their own judgement before comparing it with the system’s recommendation. That keeps alive the chance to learn how decisions are made. Giving a reviewer responsibility without the time or competence to exercise it serves neither the employee nor the customer.
At portfolio level, the fund demonstrated improvements in decisions and outcomes, and counts a decision to narrow or stop an unsuitable use as one of them. That leaves room for experiments without forcing every pilot to become a permanent programme. It also gives data work a commercial justification you can defend: the investment pays for the decisions it enables, the consequences it reduces and the capability it leaves behind.
Questions for the Boardroom
- Which recurring decision is this AI investment meant to improve, and how will we know the outcome is better?
- Where could a different definition or a missing piece of context turn a technically correct answer into a commercially wrong recommendation?
- What can each agent commit us to, and who can actually stop or correct it?
- Whose experience is missing from our records, and what would make that person willing to share uncertainty and failed decisions?
- Once preparation, review, and correction are paid for, what benefit does the business receive, and which human capability are we keeping?
Reasoning worth preserving
Go back to the purchase order, the forecast and the complaint. They are still useful records, and they become far more valuable once your organisation can explain how they relate, where they disagree and what a person or a system is allowed to do with them. AI makes it much easier to assemble an answer from such material. Deciding when that answer is good enough grounds for action remains a leadership responsibility.
The larger opportunity is to make your company’s reasoning easier to examine and improve. If you keep the uncertainty and the experience alongside the operational facts, the next person to face a decision inherits more than the last outcome. A data strategy built for that purpose gives AI something better to work with than a large pile of records: an organisation willing to explain its decisions and to learn from what they lead to.
References
The two books are discussed from summaries, not a full rereading; their bibliographic details were checked against publisher and library records. Surveys report the views of their defined samples, company accounts are the companies’ own reports, and community discussions are qualitative signals only.
- Marr, Bernard. Data Strategy: How to Profit from a World of Big Data, Analytics and the Internet of Things. Kogan Page, 2017. Bibliographic record, KIT Library.
- Schank, Roger C., Dimitris Lyras and Elliot Soloway. The Future of Decision Making: How Revolutionary Software Can Improve the Ability to Decide. Palgrave Macmillan, 2010. Publisher’s record.
- Kogan Page. Data Strategy: How to Use Data and Artificial Intelligence to Transform Your Business, third edition, published 3 September 2025.
- Fivetran. The 2026 Agentic AI Readiness Index, announcement and methodology, 5 May 2026. Survey by Redpoint Ventures of 400 data professionals at mid-sized and large enterprises.
- McKinsey. The state of AI in 2026: On the road to ROI, 25 August 2026. Survey fielded 4 May–8 June 2026; 1,719 respondents in 97 countries.
- Dillon, Eleanor, Sonia Jaffe, Nicole Immorlica and Christopher Stanton. Shifting Work Patterns with Generative AI. Working paper, April 2025. Six-month randomised field experiment with 6,000 workers.
- Reddit, r/BusinessIntelligence. What could go wrong with agent-generated dashboards, 2 April 2026. Public discussion; not representative research.
- Reddit, r/dataengineering. What is the future for dataengineering?, 23 January 2026. Public discussion; not an employment forecast.
- Simon, Eric. 2026 Data Quality Predictions: Context Graphs, AI Agents, and Transparency, LinkedIn. Practitioner predictions, not established outcomes.
- Eurostat. Digitalisation in Europe, 2026 edition. Enterprise AI figures refer to 2025 and cover enterprises with at least ten persons employed.
- BBVA. BBVA Cuts Customer Inquiry Response Times by More Than 15% in Italy and Germany Thanks to AI, updated 31 July 2026. Company-reported result.
- PKO Bank Polski. PKO Bank Polski i Visa przeprowadzają zakupy z wykorzystaniem Agenta AI, 2 July 2026. Company account of agent-initiated purchases in Visa’s Agentic Ready programme.
- Emirates NBD. Turning millions of customer conversations into better banking experiences with AI. Company description; no independent outcome data.
- Lee, Hao-Ping, and colleagues. The Impact of Generative AI on Critical Thinking: Self-Reported Reductions in Cognitive Effort and Confidence Effects From a Survey of Knowledge Workers. CHI 2025. Survey of 319 knowledge workers.
- Anthropic. Demystifying evals for AI agents, 9 January 2026. Engineering guidance on evaluation and monitoring.
- Salesforce. New Study of 2,025 Agentic AI Leaders: First To Launch Isn’t Fastest to ROI, 27 August 2026. Survey fielded 14–28 May 2026 in 20 countries; outcomes self-reported.
About the Author
Andy Demir is an international commercial and business leader with more than two decades of experience across complex B2B and industrial markets. Andy writes about executive leadership, strategy, organisational transformation, commercial performance, international business and the changing world of work, combining research with lessons drawn from operating across industries, markets and cultures.
