Skip to content
Tunisia and AI
EssayAI translation· Tunisia and AI · Part 2· 6 August 2026

The Enterprise Meets AI: Grabbing the Eel

Why impressive demos still produce so few lasting transformations, and three maxims to break the cycle.

by 15 min read

In the previous article, we set up the equation: artificial intelligence confronts Tunisia (its companies, its workers and its institutions) with several interdependent unknowns. Let’s start with the first one: the enterprise.

AI is often treated there as just another tool. Hand out a few licences, organise a prompting workshop, then wait for productivity to show up.

But the pyromaniac interior designer from the first article doesn’t stop at rearranging the furniture. He goes after the load-bearing walls: processes, responsibilities, data, software and, sometimes, business models.

If LinkedIn or the tech press are to be believed, this transformation is already done. Every week brings a new tool (or “version X” of the last one) promising to multiply our productivity tenfold, eliminate the drudge work and, while it’s at it, write a poem in the style of Abou El Kacem El Chebbi.

Reality inside companies is less glamorous. There, AI behaves like an eel: everyone wants to grab it, few actually manage to hold on.

The eel and the POC loop

What you find are licences distributed without a strategy, AI committees without a clear mandate and, above all, a multitude of pilots, prototypes and experiments that lead to nothing lasting.

Several recent studies, despite their methodological differences, reach the same conclusion: only a minority of companies manage to move beyond experimentation and generate significant, measurable value.

Many have entered a permanent POC loop.

A Proof of Concept, or POC, shows that a technology can work in principle, in a protected environment, with hand-picked data and few consequences if it gets things wrong.

The problem starts when the POC becomes the permanent form of adoption.

We end up with impressive prototypes, but few systems to which a company can genuinely entrust its operations. Between the demo and production, the same obstacles always stand in the way:

  • reliability of results;
  • integration with existing software;
  • data protection;
  • cost control;
  • traceability of decisions;
  • assignment of responsibility;
  • measurement of the value created.

Why does this gap persist? Partly because the tools offered to companies come from a world whose priorities are not theirs.

A technology born in another world

For a long time, enterprise IT was designed to structure data, standardise operations and cut costs. Accounting, databases, ERP and management software rested on a simple promise: the same input should produce the same output. Let’s call this world Business-IT.

From the 2010s onwards, the centre of gravity of innovation shifted to the consumer market. Smartphones, social platforms and free apps favoured rapid adoption: launch, watch how people use it, fix, repeat. This is the world of Social-IT.

These two worlds follow different logics.

Business-IT demands stability, control and compliance. It treats data as a sensitive asset, sometimes even as a legal liability, that must be protected.

Social-IT favours experimentation, user experience and speed of adoption. Data becomes an economic resource, and users often contribute to product development simply by using it.

Generative AI largely comes from this second world. It is versatile, instantly accessible and designed to impress from the very first use.

But companies ask it to play by the rules of the first: guarantee a level of quality, protect information, explain decisions and integrate into systems that cannot afford to improvise.

A chatbot can get something wrong in a conversation without causing a disaster. An accounting, medical, industrial or banking system enjoys no such leniency.

This mismatch explains part of the POC loop. What works in a creative demo does not necessarily meet the demands of production.

A company cannot build its critical infrastructure on the principle: ship it and see what happens.

To escape this loop, three maxims can serve as a compass.

Maxim 1: It’s always the process, stupid!

In 1992, strategist James Carville summed up the priority of Bill Clinton’s campaign in a phrase that became famous: It’s the economy, stupid!

Let’s transpose it to the enterprise:

An AI strategy that doesn’t start from processes isn’t a strategy: it’s a tool catalogue.

The economic question is always the same:

Which process do we want to improve, why, and by what measure of success?

AI doesn’t automatically bring intelligence

Like the major technologies before it, AI does not spontaneously make an organisation smarter.

It can make a process faster, cheaper, more consistent or able to handle more cases. But it also acts as an amplifier: it magnifies the effects of a good process just as much as those of a broken one.

Sprinkling “AI fairy dust” over a badly designed organisation doesn’t turn its dysfunctions into gold. Sometimes it just lets you produce the same mistakes faster and at greater scale.

So the first step isn’t choosing a model. It’s going back to the process:

1. What economic problem do we want to solve? 2. What information goes into the process? 3. What transformations and decisions does it produce? 4. Where are the errors, the delays and the repetitions? 5. Which decisions require human judgement? 6. What outcome do we expect, and how will it be measured?

That last question is essential. A demo can be “impressive”. A process has to be measurable.

Two revealing experiments

Bridgewater Associates, one of the largest investment funds in the world, did not ask an AI to “understand finance”. The firm defined a far more precise task: ranking financial documents and information by their relevance to its analysts.

General-purpose models initially delivered mediocre results. More detailed instructions improved them, but their performance eventually plateaued. A specialised model, trained on the decisions of Bridgewater’s experts, then produced roughly 30% fewer errors than the best general-purpose model tested, at a significantly lower cost of use.

Another experiment at Shopify focused on managing an e-commerce catalogue: product categorisation, brand verification, attribute extraction and application of business rules.

Here too, the problem was not framed as “using AI in e-commerce”. It had been broken down into a series of verifiable decisions. A relatively modest specialised model scored 87.3%, against 76.9% for the best general-purpose configuration tested, at a reported cost that was significantly lower.

These results are not a universal law. But they do show what separates an operational project from a demo: a precise process, a measurable outcome and a correction loop.

Success didn’t come from a magic AI bolted onto the organisation. It came from a better definition of the process.

The untapped seam of unstructured processes

That said, AI does open new territory for process optimisation: it can interpret and transform, with great flexibility, information that doesn’t naturally fit into the rows and columns of classic Business-IT.

For decades, software mostly demanded structured data: fixed fields, forms, categories and explicit rules. But a large part of corporate life happens elsewhere: in emails, PDFs, slide decks, images, contracts and customer conversations.

Since traditional systems handled this mess poorly, humans (the Knowledge Workers) served as the interface. They read, interpreted, rephrased and keyed the information into the company’s software.

AI can now take over part of these cognitive operations. But this possibility also exposes a problem:

Processes built on unstructured data are often themselves poorly structured.

“Handling a complex complaint”, “preparing a tailored offer” or “analysing an atypical case” don’t always refer to documented processes. These tasks often live in the heads of experienced staff, with their shortcuts, their exceptions and their intuitive calls.

And you can’t properly automate what you haven’t first understood.

So AI doesn’t replace design work. If anything, it makes it indispensable.

Maxim 2: My data, my AI

A company accumulates a history: decisions, mistakes, methods, customer relationships, professional practices and compromises worked out over the years.

This institutional knowledge makes up a large part of its competitive advantage. It lives in procedures and document repositories, but also in managers’ daily decisions, especially in those edge cases where the formal rule no longer suffices.

Yet general-purpose AI models wear a fundamental set of blinkers:

AI knows the internet. It doesn’t know your enterprise.

They may know the general principles of an industry without knowing why a company accepts a particular risk, rejects a particular supplier or treats a particular client differently. They know the world’s average, but not necessarily the exception that defines your business.

So the challenge isn’t just hooking a chatbot up to your documents. It’s turning institutional knowledge into an explicit, governed and assessable asset.

This is where the second maxim meets the first: the most valuable knowledge is often locked inside processes that nobody has fully documented.

The “prompt tax”

When a company uses a general-purpose model, it has to feed it the necessary context: rules, criteria, examples, documents, exceptions and expected format.

It then pays what we might call a prompt tax.

The system has no lasting knowledge of the organisation. So you have to “re-explain” your business every single time: token after token, request after request.

In the e-commerce catalogue experiment, detailed instructions increased the cost of each call by 28 to 55%. Even so, they still couldn’t capture all of the experts’ implicit decisions.

Rented intelligence is paid for on every call. Owned intelligence requires an upfront investment, but can then be reused at scale.

That’s exactly what the two experiments above illustrate. Their advantage didn’t come only from choosing a different model, but from specific data: classifications, corrections and judgements produced by experts.

The specialised model wasn’t simply learning general knowledge. It was learning how the company made one specific category of decisions.

The quality of this data matters more than its sheer quantity. At Bridgewater, classifications made by non-specialists turned out to be of little use. The experts had to step in on the contentious cases, precisely where their judgement added real value.

So the real asset wasn’t the raw document. It was the expert decision attached to the document.

Options, not a single recipe

This doesn’t mean every Tunisian company should immediately train its own model. Several paths are possible.

Renting a general-purpose model suits exploratory, creative or infrequent uses. This phase can also serve to collect the examples, corrections and evaluations needed to better understand the need.

Adapting a model in a controlled environment lets you leverage company data without having to build the entire technical infrastructure in-house. This route offers more specialisation, at the price of partial dependence on the chosen provider.

Building or training a specialised model becomes relevant when volumes are high, results verifiable, data sensitive, errors costly or the decision strategically important.

In practice, many architectures will be hybrid: a general-purpose model can coordinate, explain or handle new situations, while specialised models take care of the repetitive decisions where company knowledge matters most.

The main obstacle won’t necessarily be computing power. More often, it will remain the quality of the data and the time experts have to devote to producing it.

“Intelligence Ownership”: a question of economic sovereignty

For Tunisian companies, this goes well beyond the abstract protection of data.

An organisation that continuously hands its context over to foreign services pays for every call while outsourcing part of its cognitive infrastructure. It doesn’t necessarily lose legal ownership of its data, but it becomes dependent on a platform to turn that data into decisions.

This dependence can turn into a silent haemorrhage of know-how.

But other paths are possible.

If relatively modest models can excel at specialised tasks, the advantage no longer depends solely on the size of your data centres. It also depends on the quality of your domain knowledge.

And Tunisian companies hold valuable knowledge in finance, industry, healthcare, tourism, administrative services and in dealing with European, African and Arab markets.

So the strategic question isn’t just:

“Which model should we use?”

But rather:

“Which decisions must we continue to own?”

Maxim 3: My process, my software

While the general public debates whether AI can write poetry, a quieter transformation is reaching the backbone of digital society: software development.

Code offers models a particularly favourable environment. It can be executed, tested and fixed. Today’s tools no longer just autocomplete a few lines: they can produce components, modify an application and build tools from a need expressed in natural language.

This shift doesn’t eliminate architecture or expertise. It can, however, shorten the time needed to turn a precisely defined need into working software.

From need to software

I experienced this first-hand while developing a custom assessment app for my students.

My role wasn’t that of a traditional coder, but of a project manager: defining objectives, describing flows, reviewing results and correcting course. Agents took care of a large share of the technical production. A first working version of the app was available within a few hours.

This experience doesn’t prove that anyone can now build any system. It shows that the boundary between whoever understands the need and whoever produces the software is becoming more porous.

And that shift could have a significant consequence for companies.

The partial reversal of build versus buy

For decades, custom development was expensive. So companies bought standardised software, then forced their processes into the categories the vendor had provided.

The tool didn’t adapt to the organisation; the organisation adapted to the tool.

Some of today’s bad processes stem from this compromise. Teams built their practices around the limitations of an ERP, a CRM or a piece of software bought twenty years earlier.

When the cost of producing code falls, the equation changes. A company can more easily:

  • extend a standard software package;
  • automate a missing interface;
  • build a targeted internal tool;
  • gradually replace an overly rigid component;
  • develop an application tailored to a specific process.

Hence the third maxim:

My process, my software: software can once again adapt to the process, rather than the other way round.

The e-commerce catalogue case indirectly illustrates this shift. The team didn’t just query a chatbot. It built an environment replicating the real process: products, rules, decisions and evaluation criteria. This digital equivalent allowed the system to practise, receive a score and improve.

In other words, the process first had to be turned into software before artificial intelligence could be effectively integrated into it.

This approach opens up opportunities for Tunisian companies, particularly where international software poorly understands their constraints, their languages, their business practices or their regulatory environment.

It doesn’t mean every SME should rebuild its accounting system in a garage.

The cost of producing code is falling, but the cost of understanding the problem, integration, security, testing and maintenance doesn’t vanish. Software produced faster can also become technical debt produced faster: yet another productivity gain we’d happily do without.

Still, the company gains more options. The choice is no longer limited to buying an oversized suite or funding a lengthy custom project. More modular architectures become possible, combining:

  • standard products for common functions;
  • custom-built interfaces;
  • components generated with the help of AI;
  • specialised models for certain decisions;
  • human oversight mechanisms.

The sequence remains essential:

1. understand the problem; 2. redesign the process; 3. define responsibilities and guardrails; 4. determine which data and decisions must stay under control; 5. choose what should be bought, adapted or built; 6. use AI to speed up implementation.

Starting by generating the software would, once again, mean automating a process that nobody took the time to think through.

For Tunisian companies: choose rather than endure

The three maxims don’t describe three separate projects. They form a single sequence:

Understand the process. Own the knowledge that sets it apart (Intelligence Ownership). Build the architecture that lets you run it.

The process shows where the value lies.

Data and expert decisions make up the institutional knowledge.

Software turns that knowledge into operational capability.

So AI isn’t a magic layer you add at the end. It becomes one component of an organisation redesigned around its decisions, its information flows and its responsibilities.

A worthwhile project usually starts with a clearly identified economic problem, a verifiable outcome, available data and a success metric defined before any experimentation.

Conversely, a project framed as “we need to do something with AI” already has all the makings of a future abandoned POC.

Not all Tunisian companies will follow the same path. Some will mainly use general-purpose services. Others will adapt models to their own data. A few may turn their expertise into exportable solutions.

The point isn’t to build everything locally or to reject international platforms. Autarkic sovereignty would be as costly as it is illusory.

The point is rather to choose deliberately:

  • which technologies can be rented;
  • which data can be entrusted to a third party;
  • which decisions must stay under control;
  • which skills must be developed locally;
  • which processes can become exportable products or services.

AI can deepen Tunisia’s technological dependence. It can also help its companies turn their domain knowledge into software, specialised models and new services.

Both scenarios will probably coexist. The difference will be decided less by access to models, which are increasingly widely available, than by the ability to understand processes, organise knowledge and move from experimentation to production.

The eel isn’t getting any less slippery. But we’re beginning to understand where to grab it.

The next unknown: what becomes of the knowledge worker?

Transforming processes, owning data and producing software more easily doesn’t just change the architecture of the enterprise. It also redistributes the work.

When the machine handles the simple cases, prepares the analysis or produces a first decision, professionals increasingly have to define, verify, correct and arbitrate.

But a difficulty emerges: if AI performs the very tasks through which beginners used to learn their trade, how do we train tomorrow’s experts?

Other questions follow. How do we recognise the contribution of employees whose decisions become training data? How do we pass on knowledge? How do we organise careers when a company can produce more with fewer people?

That will be the subject of the next article: no longer what AI does to the company’s processes, but what the transformed company does to the Knowledge Worker.

Because behind every process made more efficient lies another unknown: how much room we decide to leave in it for humans.

Live long and prosper 😉✋

Dossier

Tunisia and AI

Challenges and opportunities: a series in four parts, published in collaboration with Leaders Magazine.

Part 2 · 2 by 4

  1. 1.Tunisia and AI: an equation with several unknowns
  2. 2.The Enterprise Meets AI: Grabbing the Eel
  3. 3.The knowledge worker and AI: become a centaur or vanish quietly?
  4. 4.The Governance Puzzle
Read on
Topic
Tunisia
▶Listen