An honest account of enterprise AI — and why BME does it differently

SECTION 01

The AI Implementation Crisis: Research Findings

Across the most authoritative research organisations in enterprise technology, the findings are consistent: the majority of AI implementation projects fail to deliver meaningful business value.

Source	Finding
MIT NANDA Initiative, 2025	Only 5% of enterprise AI pilot programmes achieve measurable revenue impact. The vast majority stall with little to no effect on profit and loss.
RAND Corporation, 2024–2025	Over 80% of AI projects fail to deliver their intended business value — twice the failure rate of non-AI technology projects.
S&P Global, 2025	42% of companies abandoned most of their AI initiatives in 2025, up from just 17% the previous year. The average organisation scrapped 46% of AI proof-of-concepts before reaching production.
BCG, 2024–2025	74% of companies showed no tangible value from AI investments despite $252 billion in collective spending. By 2025, 60% still generated no material value despite continued investment.
Gartner, 2024–2025	Only 48% of AI projects make it into production. At least 30% of generative AI projects are abandoned after proof of concept due to poor data quality, escalating costs or unclear business value.

The most commonly cited causes of failure are not technical. They are organisational — poor data quality, unclear business objectives, tools bolted onto existing processes rather than designed around them, and an absence of governance. All of these risks are amplified when AI is embedded deeply into a system’s core workflows.

SECTION 02

What the Largest Vendors Have Experienced

The difficulties of embedding AI into established business software are not hypothetical — they have played out publicly at the very largest vendors in the ERP and CRM market.

Microsoft Copilot

Microsoft’s Copilot has faced persistent criticism since its launch. Performance problems, hallucinated outputs, and low adoption rates have been widely reported. Independent analysts awarded it a D− rating in customer experience reviews, and Microsoft’s own CEO was reported to have described some integrations internally as approaching unusable. Despite continued investment, enterprise adoption has remained below expectations, with customers citing data privacy concerns, inconsistent reliability, and the burden of verifying AI outputs.

Salesforce Agentforce

After replacing a significant portion of its customer support function with AI and publicly celebrating the decision, Salesforce was forced to reverse course following a wave of customer complaints. Users described the experience as infuriating and called for the return of human support — a turnaround that illustrated how difficult it is to embed AI into customer-facing workflows without degrading the experience.

SAP Joule

SAP’s Joule AI assistant has seen slow adoption even among its own customer base. Six out of ten companies undertaking S/4HANA migrations are not activating Joule — not because the technology is absent, but because the scale of the migration itself leaves little capacity for additional AI transformation simultaneously.

Oracle Fusion and NetSuite

Oracle NetSuite’s own EVP set the ambition for its AI features at simply making them “correct enough of the time that you don’t want to turn them off” — a deliberately modest target that reveals how difficult reliable embedded AI genuinely is. NetSuite used an ‘undo’ button as a direct proxy for user dissatisfaction with AI outputs during testing.

IFS

IFS was acknowledged by independent analysts to have arrived late to generative AI compared with competitors. At IFS’s own customer conference in 2025, attendees noted that deriving value from AI was not about the AI agent itself — it was about everything the agent is built upon. AI capability, however well marketed, does not substitute for the foundational work that makes any system trustworthy.

Workday

A collective action lawsuit — Mobley v. Workday — alleges that Workday’s AI-based applicant screening system systematically discriminated against candidates aged 40 and over. Workday’s own court filings confirmed that over one billion applications were processed and rejected using its software during the relevant period. Separately, Workday’s own research found that 40% of the time saved through AI tools is lost again fixing AI-generated errors.

ServiceNow

AI features are available only to cloud customers — on-premise deployments are excluded entirely. Pricing is not published, with customers required to commit to long enterprise contracts before understanding their costs. ServiceNow’s own community documentation acknowledges that hallucinations are a possibility with any generative AI product.

The pattern extends across the whole market

The same pattern is visible across the entire enterprise software market, including vendors much closer in scale to the businesses BME serves.

Sage has followed the same embedded AI path with its Sage Copilot product. Sage’s own marketing materials cite McKinsey research showing that while 62% of businesses are experimenting with AI agents, only 23% have managed to scale them successfully in even a single area — an honest description of how far the industry still is from delivering on its promises.

Epicor, which serves manufacturing, distribution, and building supply businesses, launched what it describes as “the industry’s first ERP AI agent with outcomes-based pricing.” This is the consumption-based billing shift described in Section 3, dressed in different language. Customers evaluating Epicor’s AI features should understand that the pricing model for those features is explicitly variable and outcome-dependent — not included in the licence they currently pay.

Infor, serving aerospace, defence, manufacturing, and the public sector, has made a strategic decision that reveals the direction of travel: AI capabilities will be delivered cloud-first. Existing on-premise customers are explicitly told that new AI features will arrive in cloud versions first, with on-premise support deprioritised. For customers whose reasons for being on-premise have not changed, this is a statement that their vendor’s priorities have diverged from their operational needs.

The hardware problem nobody is talking about

When AI is embedded into the core of an ERP system, the compute requirements change fundamentally. Running large language models and the retrieval-augmented generation infrastructure that most vendors use requires significant GPU-based processing capacity — categorically different from the CPU and memory requirements of traditional on-premise ERP servers.

For on-premise customers, this creates a dilemma with no comfortable resolution. Their existing server hardware is almost certainly inadequate to run embedded AI locally — meaning AI features are either unavailable, degraded, or dependent on a cloud connection that routes their data outside their controlled environment. This directly contradicts the reason many on-premise customers chose that deployment model: data sovereignty, network isolation, security clearance requirements, or regulatory compliance.

The alternative — upgrading to GPU-capable servers able to support AI inference — is not a routine hardware refresh. It is a substantial capital investment requiring specialist equipment, increased power consumption, and ongoing maintenance expertise that most mid-market businesses do not have in-house. Infor’s cloud-first AI policy is therefore not simply a development priority statement. It is an acknowledgement that meaningful embedded AI on conventional on-premise hardware is not currently viable — and that the cost falls entirely on the customer.

SECTION 03

The Hidden Costs of Embedded AI

Beyond the headline failures, the economics of embedded AI are fundamentally broken — and the costs are being quietly transferred to customers who cannot anticipate them.

3.1 — Performance and compute overhead

Retrieval-Augmented Generation (RAG) — the architecture most vendors use — requires constant background indexing of database records. This increases storage and background processing load by an estimated 20–35% compared to a non-AI system. To prevent AI from blocking the user interface, most vendors have moved AI tasks to a separate processing queue. The user asks a question and must wait. The ‘two-click’ speed of a well-built deterministic system becomes a direct and unfavourable comparison.

3.2 — The billing model shift

Every time a user asks an embedded AI to reason across data, it triggers a compute bill for the vendor. On a flat per-seat licence, that bill comes directly out of the vendor’s margin. To survive, vendors are shifting to hybrid billing: a base subscription fee plus consumption credits for AI tasks. Most customers signing contracts today are unaware their monthly bill will look very different in twelve months. Agentic AI systems multiply inference costs by three to five times per task. Gartner reported some enterprises saw cloud bills spike by 40% in late 2025 before implementing guardrails.

3.3 — The customisation crisis

In an AI-embedded system, the AI frequently has no knowledge that custom business logic exists — it reasons from general patterns, not specific rules. Building a truly custom AI model for a single customer is currently priced between $150,000 and $500,000. Most mid-market businesses cannot access this. And even those that can face a further problem: every base model update — which happens frequently and without warning — may break those customisations. Research indicates legacy maintenance costs are rising by 18–25% as vendors deprioritise non-AI-compatible code.

3.4 — Legal liability and auditability collapse

Most major vendors’ terms of service now explicitly state that the customer is responsible for verifying every AI-generated output. If an AI agent reconciles an invoice incorrectly and that figure flows into a tax filing, the vendor is not liable. The customer is. This is written into the contracts customers are already signing.

3.5 — Data toxicity

When AI-generated outputs are fed back into the underlying database, errors compound over time. MIT research identifies automated data contamination as a leading cause of ERP system failure. Unlike a human error, which is typically isolated, an AI error that enters the database can propagate silently across months of subsequent processing before it is detected.

SECTION 04

The Vendor Precipice: A No-Win Situation

For fifteen years, enterprise ERP and CRM margins ran at 80% or above. Embedding AI has changed this model fundamentally and perhaps irreversibly.

4.1 — The billing trap

If vendors absorb compute costs and maintain flat subscription pricing, they eventually become unprofitable on every AI interaction. If they switch to consumption-based billing, they risk a mass customer exodus to newer, more efficient competitors who built AI into their architecture from the ground up at lower cost. There is no obvious middle path.

4.2 — The Small Language Model counter-trend

In response, 2026 has seen a significant shift toward Small Language Models (SLMs) — domain-specific models trained to do one thing extremely well. Early results suggest cost reductions of up to 80% compared to general LLM approaches. This signals that the AI approach embedded vendors bet on is already being superseded — and that significant rebuilding lies ahead. Customers will wait for that migration before they benefit.

4.3 — The state of play in 2026

Stakeholder	The illusion	The precipice reality
The vendor	“We are an AI company now.”	“Our cloud compute bills are eating our entire R&D budget.”
The customer	“The AI will handle the boring stuff.”	“I am now paying for a Digital Employee that I have to audit manually.”
The system	“Infinite flexibility.”	“Technical debt is so high we can’t update the core code any more.”

Most customers will only recognise they are on the edge of this precipice when their first usage-based bill arrives, or when an AI-driven error causes a significant financial audit failure. Simplicity and traceability are becoming the ultimate — and increasingly rare — luxury in enterprise software.

SECTION 05

The AI Industry Financial Crisis: 2026

In early 2026, approximately $300 billion in market value was wiped from the global software sector in days. Major software stocks fell 25–40% from their peaks. The ‘SaaSpocalypse’ was not caused by a single event but by the accumulated weight of a structural problem: the cost of delivering AI features is consuming the margins that made enterprise software profitable.

Gartner estimates that 60% of software spending growth in 2026 will simply cover price increases on existing software to recover AI compute costs — not fund new innovation. Nearly 70% of software providers have admitted AI delivery costs are actively eating into profitability. Nearly three quarters of organisations are actively considering switching vendors before 2028.

SECTION 06

BME’s Own Real-World Test

Using the BME API, we connected to three leading AI engines — the same engines enterprise vendors are embedding into their products — and provided each with an identical dataset: a real list of sales orders. We asked each engine to perform two tasks.

This finding directly shaped BME’s architecture. All calculations are performed by BME’s own deterministic query engine before any data reaches the AI. The AI never does the maths. BME does the maths. The AI explains what the numbers mean.

SECTION 07

The Core Principle: Push, Don’t Pull

• When you request an AI-generated insight, BME gathers the relevant data internally and performs all calculations
• Verified results are assembled into a structured package and sent to your chosen AI provider in a single, focused request
• The AI receives pre-calculated, correct figures and produces a plain-English narrative
• The data is not retained by the AI provider beyond the scope of that single request

Your ERP remains the single source of truth. The AI is a reasoning engine — not a calculator, and not a second database.

SECTION 08

BME Calculates. The AI Explains.

The most important design decision in BME Pulse is the simplest: BME calculates. The AI explains. This is the conclusion the enterprise software industry has reached after two years of expensive lessons — and the architecture BME adopted from the outset.

When a BME Pulse response says a customer spent £38,420 in Q1, that figure was calculated by BME — the same figure that would appear in a printed report, a list view, or an export. The AI did not produce it, estimate it, or approximate it. It simply reported it in plain English, with context.

SECTION 09

How It Works: Six Steps

Every BME Pulse interaction follows the same clean pattern. BME does the data work. The AI does the reporting.

Step	What happens	Who does it
1	You ask a question in plain English — on a Sales Order, inventory record, or list view	You
2	BME runs queries against your database and completes all mathematical operations internally. Nothing has left your system	BME
3	Pre-calculated, verified results are packaged with a clear instruction: report on these figures, do not calculate	BME
4	The verified package is sent to your chosen AI provider using your own API key	BME
5	The AI produces a plain-English report — explaining patterns, highlighting what matters, adding context	AI
6	No data retained externally. No record held. The next query starts completely fresh	—

SECTION 10

Bring Your Own Key (BYOK)

BME does not act as an intermediary between you and your AI provider. Instead, you connect directly:

• You create your own account with your chosen AI provider (Anthropic, OpenAI, DeepSeek, or others)
• You obtain an API key from that provider and enter it once in BME’s AI settings
• Your AI usage is billed directly from the provider to you — BME does not mark up or resell AI capacity
• You can monitor your own usage and costs through your provider’s own dashboard
• You can change or switch AI provider at any time without affecting your BME data or workflows

BME charges a module fee for BME Pulse, covering development and ongoing maintenance. The cost of running AI queries is entirely your own, at the rates set by your chosen provider — typically a fraction of a penny per query.

SECTION 11

Custom AI Queries

Where a customer needs specific analytical capability that goes beyond standard list views, BME will build the necessary query to make it possible. Where analysis spans multiple related datasets, BME builds a custom internal query, performs all calculations, and passes verified results to the AI — without any compromise to data security or the push philosophy.

Custom AI queries are scoped, built, and tested as a standard BME development engagement. Once in place, they are available permanently, can be refined over time, and work with any AI provider the customer chooses.

SECTION 12

BME Pulse: The AI Module

BME Pulse is a single, growing product that brings AI into your ERP workflows on your terms. It launches with a focused core feature set and will expand continuously. Every capability your chosen AI partner develops is immediately available to you.

BME Pulse operates entirely within your active BME session. When your session ends, Pulse goes with it — displaying a clean “Session disconnected” state. AI activity only occurs when you are present and authenticated.

SECTION 13

Resilience: Your Business Doesn’t Stop

Because BME’s AI layer sits entirely outside the core ERP, an AI provider outage is an inconvenience rather than a crisis.

Scenario	BME approach	Embedded AI
AI provider outage	BME continues working fully. AI insights temporarily unavailable. No core function affected.	Features dependent on AI stop working. Core workflows may be significantly impacted.
Slow AI response	Query takes longer. The rest of BME is completely unaffected.	UI delays or timeouts may affect core application screens.
AI price increase	Customer negotiates directly or switches immediately. BME is unaffected.	Vendor absorbs or passes on the increase. Customer has no visibility or control.
Better AI model released	Customer switches with a single setting change. Immediate benefit.	Customer waits for the vendor to re-engineer and release an update.
AI provider discontinued	Customer moves to an alternative freely. No BME change required.	Vendor must rebuild. Customer is blocked until the work is complete.

SECTION 14

Future-Proofing: Never Locked In

• A better model is released — switch to it immediately, often with a single setting change
• A cheaper model emerges — move to it and costs fall directly
• Your current provider raises prices — evaluate alternatives without raising a support ticket with BME
• A provider exits the market — switch, and BME continues without any change required

SECTION 15

BME’s Structural Advantages

BME’s architecture, licensing model, and philosophy address almost every risk that the current crisis has exposed.

Runs without AI

Every report, process, calculation, and workflow operates completely independently of any AI connection. If AI is unavailable, nothing in BME stops working. AI is an enhancement, not a dependency.

Runs in complete isolation

BME can be deployed entirely on the customer’s own hardware, on their own network, with no external connectivity. No cloud dependency, no third-party infrastructure, no exposure to external outages or pricing changes.

Perpetual licensing

Customers own their software. No subscription that can be restructured, no usage-based component that escalates unpredictably, no vendor decision that changes what the customer pays. In a market where SaaS pricing is becoming a source of significant financial risk, perpetual licensing is a genuine and increasingly rare advantage.

Concurrent licensing

Pay for active users, not registered users. For businesses with shift patterns, seasonal peaks, or varying usage across departments, this delivers substantial cost efficiency that per-seat SaaS models cannot match.

No AI cost volatility

AI costs are billed directly from your provider to you. BME has no exposure to compute cost volatility, and customers can monitor, cap, and control their AI spend independently.

Architecture that ages well

By keeping AI outside the core ERP, BME avoids the AI technical debt, the upgrade tax, and the risk of base model updates breaking established customisations that now burden every vendor who chose the embedded path.

SECTION 16

Advantages and Considerations

Advantages of BME’s push philosophy
Deterministic accuracy	BME performs all calculations before passing results to the AI. The AI never does arithmetic — returning the same correct answer every time.
Data sovereignty	Your data is never held externally. Sent on demand, used once, not retained.
No additional security layer	Data flows outward on request. No complex security boundary is needed between BME and an external AI service.
Regulatory compliance	Data remains in your control at all times, significantly simplifying GDPR and similar obligations.
Resilience	BME continues operating normally if your AI provider has an outage. No core function is affected.
Provider freedom	Connect any compatible provider. Switch at any time with no changes to BME, your data, or your workflows.
Immediate AI improvements	When a better or cheaper model is released, switch immediately without waiting for a vendor update.
Direct cost transparency	AI costs billed directly from provider to you. BME does not mark up or resell AI capacity.
Honest, contextual responses	Responses carry caveats, risk flags, and reasoning — not just a bare number.
Non-invasive architecture	AI sits entirely outside the core ERP. System stability is never at risk from AI behaviour.
Modular and optional	BME Pulse is opt-in. Customers who do not need it simply do not activate it.

Considerations specific to this philosophy
API account setup required	Customers must create their own account with a chosen AI provider. A one-time step, but an additional task embedded approaches would not require.
Proactive monitoring requires configuration	Session-based monitoring requires BME Pulse to be configured with the relevant query set. Data is only ever sent to AI when a user is present and active.
Each query begins without context	The AI has no memory of previous interactions. Every query is answered independently from the data provided in that moment.
Query scope is defined in advance	The data in each query package is determined at design time. An embedded AI could retrieve additional context autonomously. BME’s approach cannot.

SECTION 17

A Note on Data Currency

Every AI-generated insight in BME Pulse reflects the state of your data at the moment the query is run. Inventory levels, staff availability, build queue depth, and supplier lead times can all change between asking a question and acting on the answer.

This is not a limitation unique to AI. Any ERP calculation carries exactly the same caveat. What BME Pulse adds is transparency: rather than returning a bare number with no context, responses carry caveats, confidence notes, and risk flags. For example:

SECTION 18

Summary

BME’s approach to AI — and to software more broadly — is built on seven commitments:

• Your data stays yours. BME never stores your business data with an AI provider. Data is sent on demand, calculated by BME, and used for that query only.
• The numbers are always right. BME performs all calculations. Every figure in a BME Pulse response comes from BME’s own deterministic query engine.
• You choose your AI partner. Compatible with the leading AI providers. Never locked to any single vendor. You manage your own costs directly.
• Your system keeps working. BME is never dependent on an AI provider’s uptime. AI is an enhancement, not a foundation.
• You benefit from AI progress immediately. When better or cheaper AI models emerge, switch without waiting for a vendor update.
• Your costs are predictable. Perpetual and concurrent licensing means no subscription restructuring, no usage-based escalation, no compute cost volatility.
• Your infrastructure is your own. BME can run on your own hardware, on your own network, in complete isolation if required.