DocumentationLogin
Enterspeed logo
Enterspeed Blog
Thoughts & Insights

10 things about controlling AI and its hallucinations

Toke Lund
Toke Lund
CEO at Enterspeed
Thumbnail for blog post: 10 things about controlling AI and its hallucinations
Intro

“How do we trust AI?”

The question I hear most often these days.

Every retailer I've met that uses large language models has seen hallucinated output: a product description that invents a specification, campaign text that contradicts brand guidelines, or customer-facing copy with the wrong price. Errors like these show up in every enterprise that scales AI, which makes hallucinations an architecture question.

AI hallucinations come from how language models work: they predict the most statistically likely next word rather than the most factually correct one. What you can control is what the model predicts from. Controlling hallucinations therefore means designing the environment AI works in, above all the context it receives.

Weak context is also why most AI projects stall after the pilot. The Context Factory is how we fix that. It turns the data you already have into precise context, in three steps: data in, context out, AI at scale. This paper covers ten things that work when you deploy AI at enterprise scale, written for the people who have to make it work.

Toke Lund
Toke Lund
CEO at Enterspeed
Toke Lund, CEO of Enterspeed

First published April 2026. Updated November 2026: this edition adds Enterspeed's own benchmarks and results from XL-BYG, the Danish builders' merchant that checks its product content with Speedtrain.

Prefer to read it as a PDF?The full white paper, designed for print and sharing

Download the PDF

Not all context is equal

Give AI the right data at the right time: prepared in advance, served to match each question.

The most effective way to reduce hallucinations is to give AI the right data at the right time. Enterprises often either flood models with irrelevant information (creating noise) or starve them of necessary context (forcing the model to guess).

For a retailer, this plays out every day. Your AI writes product descriptions from a data model that hasn't been updated since last season. Or it sees everything (deprecated SKUs, old supplier copy, conflicting specifications) and doesn't know which source to trust.

The fix is better-prepared data. Your AI needs structured, validated and context-rich data, prepared in advance and served to match each question. For retail, that means product data, pricing, stock and content read from one governed truth over the systems you already run. Your ERP and PIM stay authoritative. Feeds too: Google Shopping and each marketplace need their own shape, built from the same prepared data as the product page.

Preparing the data up front pays off twice. The AI reads less, so each answer costs less and arrives faster. And it reads from one place, so fewer things can fail along the way.

Enterspeed benchmarks
Cost and speed

The same enrichment task, run on raw product data and then with the product taxonomy prepared up front.

300,00015,000

tokens used

5×cheaper
3×faster
Samequality
A token is the unit of text a model reads and is billed for.
Reliability

How often an AI agent's look-ups work, out of 100, as it fetches live from more systems.

05010013579109244failed look-upsSystems AI fetches from live
Enterspeed's own measurement. Method available on request.

When the AI answers from one prepared version of your authoritative data, it has far less to guess.

Human-on-the-loop: oversight that scales

Let automatic checks flag the risky outputs, so people review the exceptions instead of everything.

You can't have a human approve every AI output when you're processing thousands of product descriptions, campaign texts or customer interactions daily. But abandoning oversight entirely isn't the answer either.

The right model is "human-on-the-loop": intelligent flagging systems that identify which outputs need human attention and let the rest flow through. Think of it like quality control in manufacturing. You don't inspect every unit. You build processes that catch anomalies and flag edge cases.

In practice, this means confidence scoring: the system knows when a model is uncertain, outside its training domain or handling high-stakes outputs, like compliance-sensitive product claims.

You can only act as a human-on-the-loop if your data carries the metadata to flag risks automatically. Without proper data architecture, you're back to manual review or no review at all. That takes tagging, structuring and enriching your data so the system knows what's sensitive, what's verified and what needs a human eye.

In practice: XL-BYG

XL-BYG, the Danish builders' merchant, runs its product content this way. Speedtrain from Enterspeed generates and checks each product text. Only the texts that fail a check go to a product specialist, and in XL-BYG's setup 99 percent pass automatically.

30 min2 minper product
"Our product specialists validate rather than write. That uses their expertise better."
Kim Portz Knudsen, Head of Webshop, XL-BYG
The 99 percent reflects XL-BYG's workflow and data quality. A typical first run approves around 60 percent automatically.

Your training data is only half the battle

Training is a snapshot. Retrieval keeps AI current with today's ranges, prices and rules.

Training data quality matters, but production AI systems also have to stay current. This is particularly acute in retail, where product ranges, pricing, promotions and compliance requirements change constantly.

Your beautifully trained model from six months ago doesn't know about this quarter's product launches, pricing changes or new compliance requirements.

Retrieval-augmented generation (RAG) solves this. Instead of relying purely on what the model "remembers" from training, you dynamically inject current, relevant information at the moment of generation. The AI then reflects your current business rather than a snapshot from months ago.

What your AI knows, with and without RAG
Training cut-offTodayWithout RAGNew rangePrice changeNew ruleUnknown to the modelWhat it learned in trainingWith RAGCurrent data, added when the AI answers
Swipe to see the full timeline

RAG is only as good as your retrieval system. If you're pulling irrelevant documents or outdated information, you've traded one problem for another. The retrieval mechanism needs to be intelligent, fast and deeply integrated with your business logic. That depends on data architecture. When a specification is corrected in the PIM, every text built on it should follow, in every language, instead of someone fixing six translations by hand.

Data governance prevents hallucinations

Governed data tells AI what it may use, and when to say “unknown” instead of guessing.

Imagine your AI assistant:

serving up product information for a line you've discontinued,
generating content from a supplier spec sheet superseded six months ago,
or using pricing data from a market you no longer serve.

These are compliance problems, and they are also organisational hallucinations, where the AI's output conflicts with your actual business reality.

Proper data governance creates boundaries that prevent AI from accessing or surfacing information it shouldn't. This means implementing real-time policy controls that understand data sensitivity, user permissions and regulatory requirements. An AI agent should see only what the person or system asking is entitled to see.

When these guardrails are built into your data, you prevent compliance violations and a whole class of hallucinations where AI makes up responses because it lacks proper context about what it can and can't reference. The data can also say what it doesn't know, so "unknown" becomes an answer the AI can give instead of a guess.

The same question, and the care instructions are missing from the product data:
Without governance
Is this jacket machine washable?
Yes, at 40°C.
A likely-sounding guess. If the jacket is dry-clean only, it gets ruined and returned.
With governance
Is this jacket machine washable?
I don't have care instructions for this jacket. Please check the label.
The data marks care instructions as unknown, so the AI says so instead of guessing.

The enterprises doing this well have moved beyond static access controls to dynamic, context-aware data policies.

Poor governance creates legal risk and AI that contradicts your own business.

Standardisation is your friend (even when it's boring)

One standard for connecting AI to data gives you one place to validate, monitor and govern.

The dirty secret of enterprise AI is that a lot of teams are rebuilding the same things: custom integrations for every data source, bespoke connectors for every tool and one-off implementations that are impossible to scale. This fragmentation wastes engineering time and creates hallucination vectors, because every integration is a potential point of failure.

Without standardisation, you're building legacy debt with every new AI integration.

Each custom connector becomes technical debt that someone has to maintain, update and eventually replace. Worse, these one-off solutions make it nearly impossible to implement consistent validation and governance across your AI systems.

Open standards like the Model Context Protocol (MCP), which defines one way for AI to connect to data and tools, are changing this dynamic. Instead of writing custom code to connect your AI to Google Drive, then Slack, then your database, then your CRM, you build against a single protocol. Think USB-C for AI: one standard that works everywhere.

ONE CUSTOM CONNECTOR PER SYSTEMAIcustomGoogle DrivecustomSlackcustomDatabasecustomCRM
ONE STANDARD FOR ALLAIMCPGoogle DriveSlackDatabaseCRM

This standardisation matters for hallucination control because it allows you to implement consistent validation, monitoring and governance across all your data sources.

When you have uniform interfaces, you can build uniform safety measures. Plus, the time you save on integration work can be redirected to what actually matters: prompt engineering, testing and validation.

Not all hallucinations are created equal

Match controls to risk: creative tasks can run loose, compliance content cannot.

There's a crucial distinction between creative hallucinations and factual hallucinations. In brainstorming, campaign ideation and product naming, some speculation is actually useful.

The problem arises when you need factual accuracy, but your system is configured for creativity. This is why temperature settings, system prompts and model selection matter so much. A customer service bot should be configured conservatively, but a marketing ideation tool can run hotter.

Smart enterprises don't deploy one-size-fits-all AI. They build task-specific configurations with controls matched to the risk, down to single fields: a compliance claim gets required wording and forbidden terms, while a campaign tagline gets rules for tone and brand.

A content brainstorm tool and a product compliance tool should not have the same guardrails.
Use case
Typical controls
Tolerance
Risk
Creative / ExploratoryIdeation, brainstorming, marketing concepts
Loose prompts, high temperature
High
Low
Informational / OperationalInternal assistants, support responses
RAG, confidence scoring, source validation
Medium
Medium
Critical / RegulatedLegal, financial, medical, compliance
Human approval, audit logs, strict governance
Near zero
High

For retail, product compliance content, legal text and pricing sit firmly in the "critical" category, and campaign ideation and draft copy in the "creative" one. Treating them the same way is where things go wrong.

Your product feed is critical too. Google Merchant Center rejects a product when a required attribute is missing, and may do so when the price doesn't match the product page, often without anyone noticing until sales drop. Get it right and the feed pays back: after enriching its product content, XL-BYG saw 12% more clicks from Google Shopping.

Monitoring and feedback loops are non-negotiable

Watch for drift, let people flag errors, and fix the cause in the data, rules or prompt.

Monitoring AI outputs means more than logging requests and responses. It means tracking hallucination patterns, measuring output quality drift over time and identifying which types of queries consistently produce unreliable results.

The best systems implement real-time telemetry that detects when model behaviour starts diverging from expected patterns, before it reaches customers or gets published. This requires building feedback mechanisms where editors, marketers and operations teams can flag incorrect outputs, which then feed into improvement cycles. Version the rules your AI works by, too, so a change can be tested against the previous version before it becomes the default.

You wouldn't deploy a new ecommerce feature and never monitor its performance. AI systems need the same discipline – arguably more, because AI behaviour can drift in ways that traditional software doesn't.

If you're not monitoring your AI outputs systematically, you're flying blind.

Explainability is a control mechanism

If you can trace why AI answered as it did, you catch errors before they compound.

When you can trace why an AI produced a particular output (which data sources it weighted, which context it prioritised), you can spot hallucinations faster.

This becomes particularly powerful when combined with domain expertise. A product manager who can see which parts of the data the model weighted heavily can quickly identify when it's over-indexing on irrelevant information.

In retail, explainability matters at every level: compliance teams need audit trails, editors need to understand why a description was generated a certain way, and leadership needs confidence that the system is operating as intended. In practice, that means recording for every output where each value came from and which rules it passed, so guesses become something you can count.

Black box versus glass box AI
Dimension
Black box AI
Glass box AI
Visibility
None
High
Data sources
Unknown
Explicit and traceable
Decision logic
Opaque
Interpretable
Error detection
Reactive
Proactive
Human oversight
Difficult
Built-in
Regulatory readiness
Low
High
Trust level
Assumed
Verifiable

When you understand why AI made a decision, you can catch errors before they compound.

The temperature setting everyone forgets

Temperature sets how creative the model is. Treat it as a business decision about risk.

Model temperature controls the creativity-accuracy trade-off. Higher temperatures make models more creative, and more likely to hallucinate. Lower temperatures make them more deterministic, which is exactly what you want when generating product specifications, compliance copy or legally required text.

Low temperature
More deterministic
High temperature
More creative, more likely to hallucinate
Product specificationsCompliance copyLegally required text
Campaign ideation

The mistake most teams make is using default settings across all use cases. Your product content tool should run conservatively: a low temperature where the model allows it, and strict rules where it does not. Your campaign ideation tool can run much higher. These aren't just technical parameters – they're decisions about how much creative latitude your AI is allowed to take with your brand.

The key is matching temperature settings to your specific use case and then testing rigorously. Find the sweet spot between useful creativity and reliable accuracy for each application, and document it as part of your AI governance framework.

Temperature isn't a technical detail. It's a business decision about acceptable risk.

Trust, but verify, at every level of the architecture

Verify at every layer: the winners catch hallucinations before they cause damage.

The fundamental principle underlying all of this: no matter how good your controls are, you need verification mechanisms baked into your architecture. At every layer.

The retailers moving fastest with AI right now aren't the ones who've found a way to trust AI blindly. They're the ones who've built layered systems that catch and correct errors before they reach a product page, a campaign or a customer. They architect for failure, because they know failure will happen.

Trust is built not by assuming correctness, but by verifying it at every level. Below, you'll see what that looks like when you translate it from principle into practice.

The enterprises succeeding with AI aren't the ones who've eliminated hallucinations. They're the ones who catch them before they cause damage.

What this means in practice

Five controls and six layers of defence that turn the ten things into a working system.

  • Automated validation against authoritative data sources for factual claims
  • Confidence scoring that surfaces uncertain outputs for review
  • Human approval workflows for high-stakes, customer-facing content
  • Audit trails so you can trace exactly how an output was generated
  • Staged rollouts where new AI capabilities are tested at limited scale before full deployment
Defence-in-depth for AI hallucination control
Read from the bottom: each layer catches what the one below lets through.
Layer
Purpose
Risk addressed
Human oversight
Final accountabilityHigh-stakes errors
Monitoring
Detect drift and patternsSilent degradation
Model configuration
Control creativityOver-speculation
Context retrieval
Ground answersKnowledge gaps
Governance and access
Enforce boundariesUnauthorised data use
Data quality
Trust the inputsGarbage-in hallucinations

Hallucination mitigation cannot rely on a single control.

Effective systems apply layered defences across data quality, governance, context injection, model configuration, monitoring and human oversight.

Each layer reduces risk, and together they create resilient, auditable AI architectures.

So, what should you do next?

Three questions to start with, the Context Factory, and an invitation to our Feed Integrity morning.

The good news: you don't need to reinvent the wheel. The tools, standards and architecture patterns exist. What's needed is the discipline to treat data quality and governance as strategic priorities, not IT hygiene.

The companies getting this right are discovering that the same infrastructure that reduces hallucinations also makes their AI faster, more accurate and more valuable. Better data architecture reduces risk and lets AI do much more for your business.

Trustworthy AI isn't about having perfect models. It's about having the right systems in place for when they're imperfect. And they will always be imperfect.

Start with three questions
01Where is your AI currently generating outputs without a reliable, real-time data source to ground it?
02Which of your AI use cases carry the highest business risk if they hallucinate, and do they have the right guardrails?
03Do you have monitoring in place that would tell you if output quality started to degrade?
The Context Factory

Data in. Context out. AI at scale.

Data in

Every source, from products, prices and stock to supplier files and PDFs, kept fresh in real time.

Context out

Connected, gaps filled, errors fixed and adapted to each market. Your team stays on the loop and only sees the exceptions.

AI at scale

The same precise context for your webshop, feeds, marketplaces and AI agents, and back to your PIM.

Join us on 9 December

Join us in Aarhus for Retail i en AI-drevet virkelighed – om Feed Integrity. A morning for retail leaders on keeping your Google Merchant Center feed accurate for search, Shopping ads and AI agents. In Danish.

Sign upOr get in touch: info@enterspeed.com · +45 21 55 18 25

Take it with you: the full white paper as a PDF.

Download the PDF

About Enterspeed

Who we are, who we work with, and how we keep your data secure.

Enterspeed is the Context Factory. We turn the data you already have into precise context, so AI makes it out of the pilot and into production.

We take in data from the systems you already run, refine it with AI under your rules, and serve it as one precise context to your webshop, your feeds and your AI agents. Your team handles the exceptions.

Trusted by enterprises including
XL-BYGCarlsbergDanish CrownHummelFlüggerWhiteawayNovicell
ISO/IEC 27001 ISMS
Enterspeed's Information Security Management System is certified to ISO/IEC 27001 by Insight Assurance, and Enterspeed is GDPR compliant.
Data in. Context out. AI at scale.
Data in. We read the systems you already run, from PIM and ERP to supplier files and PDFs, without an integration project or a migration.
Context out. We write and check product content with Speedtrain, against your own rules, and return it to the PIM you already use.
AI at scale. We serve the same precise context to your webshop, feeds, marketplaces and AI agents, and your team can ask what's wrong with the data in Claude or ChatGPT.
Control. AI agents get read-only access, so they see only what their caller is entitled to see, and your systems stay closed.
Toke Lund
Toke Lund
CEO at Enterspeed

CEO and Partner at Enterspeed. Speaker on digital transformation and the future of Enterprise tech. Loyal lover of Liverpool! ⚽❤️

Ready to try out Enterspeed? 🚀

Start connecting & combining your services today

Solutions

Product OnboardingPublish-ready product dataGoverned AI contentReplatform and migration
Enterspeed logo

© 2020 - 2026 Enterspeed A/S. All rights reserved.

Made with ❤️ and ☕ in Denmark.