You watched a peer’s company spend £180,000 on a CRM implementation in 2014. By 2017 they were on their third “rescue” project. By 2019 they were back to spreadsheets for half the use cases. You knew the pattern was preventable. Now you are being pitched on AI strategy, and the proposal looks oddly familiar.
This is the recognition that founders aged 40 to 55 keep flagging in conversations about AI. The story of the latest enterprise software cycle is not new. The CRM, ERP, and BI cycles taught SMEs exactly which costs explode and why, and AI is replaying the same shape with different vocabulary. The lessons are already paid for. They just need to be remembered before the next round of invoices arrives.
What did the CRM cycle teach SMEs in the 2010s?
CRM was the cycle most SMEs encountered first, between roughly 2010 and 2020. Salesforce, Microsoft Dynamics, Sugar, HubSpot, and a dozen smaller platforms all promised to centralise customer data and improve sales productivity. The headline pricing looked manageable. The total cost of ownership did not. By the end of the cycle, the published rule of thumb across consulting firms and analysts was that CRM TCO ran 2 to 3 times the licensing cost over a five-year horizon, with the majority of the spend going to data migration, customisation, integration, and user training.
The deciding factors for whether a CRM implementation succeeded were almost never about the software. They were about data quality going in, change management once the system landed, and post-launch support to keep the platform from sliding back into spreadsheets. Implementations that under-invested in data migration failed regardless of platform choice. Implementations that under-invested in user adoption succeeded technically and failed culturally.
What did the ERP cycle look like at SME scale?
ERP came in waves through the 2000s and 2010s, and Panorama Consulting’s annual industry report has tracked it consistently. The 2026 edition continues to show 70 to 80% of ERP implementations failing to meet their original objectives, which is a number that has barely moved in two decades. NetSuite’s published SME example shows a 5-year TCO of around $222,000, broken down as $139,000 upfront licensing, $43,000 implementation services, and $40,000 ongoing support and customisation. The implementation services line is what kept doubling against the original budget.
The pattern was consistent across vendors. Scope creep during configuration. Data integration costs that surfaced months after sign-off. Customisation that locked the SME into a specific consultant relationship. Post-launch support running 10 to 20% of implementation cost annually for the next several years. Each one was foreseeable. Each one was repeatedly under-budgeted.
What did the BI cycle teach about tools without architecture?
The Business Intelligence cycle from 2015 to 2025 added a different lesson. Tableau, Power BI, Looker, and Qlik were bought enthusiastically by SMEs hoping to “become data-driven.” Most did not. The tools were sound; the data architecture underneath them was not. Without clean source data, defined metrics, and someone whose job it was to build dashboards, the tools sat at 5 to 15% of their licensed seats actually being used. Power BI Pro at $14 a month and Tableau Creator at $75 a month per the Tech Insider 2026 comparison illustrate the order-of-magnitude pricing variation, and most of the spend went to seats nobody opened.
The lesson SMEs eventually learned was that platform selection alone did not deliver value. The infrastructure underneath, the people equipped to build with it, and the use cases driving the work all had to be in place first. Tools without architecture became a recurring cost without a corresponding output.
What is repeating in AI right now?
AI is recapitulating these patterns with remarkable fidelity. AgentMode’s 2025 analysis of 127 enterprise AI implementations found total costs running 3.3 times initial budget, with 70% of spend in hidden categories. That is the same 2-to-3x TCO rule from CRM, with slightly more variance because AI tooling itself is more expensive and the technology cycle is faster.
Data quality is the dominant obstacle, exactly as it was for CRM and ERP. Informatica’s 2025 CDO Insights survey found 43% of organisations citing data quality and readiness as the top obstacle to AI value. Precisely’s 2025 data quality study put it at 64%, up from 50% in 2023. The numbers move slightly. The conclusion is the same as it was 15 years ago: do not buy a tool before fixing the data.
Change management is under-budgeted at exactly the same ratio as in the 2010s cycles. McKinsey change-programme research puts the typical change management allocation at around 10% of digital programme budgets, where the appropriate level is 66%+. MIT Sloan’s 2025 research on AI adoption found 70% of failures traced to inadequate change management rather than technology. The lessons are not new. The under-investment is not new.
What is different about the AI cycle?
Three things are genuinely different about AI compared to the 2010s enterprise software cycles. The technology cycle is faster. New models and capabilities arrive every three to six months, which means architecture decisions made in January can be obsolete by July. The hype-to-reality gap is wider. AI is sold as autonomous, delivers as narrow assistance, and the gap creates buyer disappointment that the previous cycles did not have to manage at the same scale. The methodology is less mature. CRM and ERP eventually accumulated standard implementation playbooks; AI is still in the phase where every consultant is improvising.
These differences raise the variance of outcomes but do not change the fundamental dynamics. The 2-3x TCO rule still applies. The data dependency still applies. The change management investment still applies. The faster pace and louder marketing make it easier to forget the lessons, which is why the 2026 failure data looks so similar to the 2010 data.
The four lessons that are already paid for
Four lessons from the 2010s enterprise software cycles apply directly to AI now. Each one was paid for by SMEs who lived through it, and each one is replayable as a budget discipline rather than a learning experience.
Budget for 2 to 3 times the headline cost. The visible fee is one third of the real spend. AgentMode’s 3.3x rule for AI matches the CRM rule from 15 years ago.
Invest in data architecture before tooling. Tools sat unused in BI. The same pattern is unfolding with AI. Without clean source data and defined operational metrics, the platform layer cannot produce value regardless of how sophisticated it is.
Treat change management as half the project, not 10% of it. The McKinsey 5.3x success ratio for culture-led change programmes versus technology-led ones is the same finding the CRM analysts published in 2015. The investment level required has not changed.
Do not buy a tool before defining a use case. The most expensive sentence in any technology cycle is “we should get one of those.” Define the use case, define the success metric, then choose the tool. The reverse order is what produced the 70 to 80% failure rate in ERP and is now producing similar numbers in AI.
The founder who lived through the 2010s technology cycles already paid for these lessons in real money and real time. The same founder is the one being pitched on AI strategy in 2026. Recognising the shape of the proposal as the same shape they saw before is the cheapest possible advantage in this cycle.
If you would like to talk about how to apply the lessons of the 2010s to your AI engagement, book a conversation.
