Leaving Salesforce3 min read

Leaving Salesforce, in the order we would do it

By ·Founder, K3 LabsPublished

We moved ARC Gaming & Technologies off Salesforce in six months, December 2025 to May 2026. Build the data layer before the CRM, and plan to run both systems side by side: we did for the whole six months.

We decided to leave in December 2025. The reason was simple: ARC's sales team could not use Salesforce to sell more, and they needed a second instance for a new geography that they could not afford. They were paying for a lot and using a small fraction of it. That is our estimate, not a measured number. The full list of reasons is in the six-month account.

If we were starting again, this is the order we would follow.

  1. Read your own order form. The renewal and notice terms in it set your deadline. We do not know what your contract says, and nobody else's blog does either.
  2. Export your data and files before you cancel anything.
  3. Build the data first and the CRM second.
  4. Run both systems side by side instead of switching over in one go, and plan for the sync between them to break.[1]
  5. Cancel last, once the new system is carrying the work.

Steps three and four are the ones we can speak to from this move, so the rest of this page is about them. The longer account is in how the six months went.

We built the data layer before the CRM

We built the Single Data Backbone and automated the reporting first. The CRM came after. We think that order is why the rest held together.

The other decision was not to copy Salesforce's shape. For ARC the CRM was an operational tool, not a sales tool. So in the build we separated sales operations from the CRM.

The sync between old and new is where the time went

Both systems ran side by side for the whole six months, December to May.

After we moved the CRM, operations still needed the Accounts page in Salesforce, and the data on the two sides went out of step. We built bi-directional pipelines to keep them in sync. Keeping copies that are updated separately consistent is a known hard problem.[2][3] A plain field-to-field mapping was not enough. Data-integration researchers say the same: mappings take business knowledge of what the data means, and the same value can mean different things in different contexts.[4][5] It took an agentic judgement layer, one that understands what each change means before it writes it to the other side.

The sync fixes went on throughout the six months. They were never one sprint. Budget for that before you start.

The field app nearly kept us on Salesforce

One thing nearly kept us on Salesforce: the field iOS app. The licence for it was expensive. We almost stayed for it anyway.

We built Helix, a real-time, offline-first app, to replace the field apps. It is not adopted yet, so we do not know whether it does the job, and we will not claim it does. That is the one place Salesforce nearly won.

Who should not move

Do not move if data is not your problem, or if you do not need to manage workflows and automation.

Our view is that Salesforce is too expensive, too big and too complicated for a small team. We also think you are better off with something built for your team on a single data backbone than with spreadsheets managed through LLMs and copilots. There is a side-by-side with Salesforce plus QuickBooks, and ARC's story shows what we built for them.

If you are deciding, talk through your exit with us. It is 30 minutes and it is a working session, not a pitch.

Our own research

This is ours, not published research. Each item says how far it goes.

  • Observation. At one customer we built the data layer before the CRM and ran both systems side by side for six months. The sync between them needed fixing throughout. One customer, one team, one industry.
  • Hypothesis. Building the data layer first reduces rework later. We built it first once and cannot say what the other order would have cost. To test it we would compare the rework of a data-first move with a CRM-first move on similar accounts.
  • Hypothesis. A judgement layer that works out what a change means handles mismatches that fixed field mapping misses. We have seen it work at one customer and have not compared it with a rules-only approach on the same changes.

References

  1. 1.Bisbal, J., Lawless, D., Richardson, R., O'Sullivan, D., Wu, B., Grimson, J., & Wade, V. (1997). A survey of research into legacy system migration. Technical report TCD-CS-1997-01, Trinity College Dublin. https://publications.scss.tcd.ie/tech-reports/reports.97/TCD-CS-1997-01.pdf
  2. 2.Saito, Y., & Shapiro, M. (2005). Optimistic replication. ACM Computing Surveys, 37(1), 42-81. https://doi.org/10.1145/1057977.1057980
  3. 3.Czarnecki, K., Foster, J. N., Hu, Z., Lämmel, R., Schürr, A., & Terwilliger, J. F. (2009). Bidirectional transformations: a cross-discipline perspective. GRACE meeting notes, state of the art, and outlook. In Theory and Practice of Model Transformations (ICMT 2009), Lecture Notes in Computer Science, pp. 260-283. Springer. https://doi.org/10.1007/978-3-642-02408-5_19
  4. 4.Halevy, A., Rajaraman, A., & Ordille, J. (2006). Data integration: the teenage years. Proceedings of the 32nd International Conference on Very Large Data Bases (VLDB '06), Seoul, Korea. https://www.vldb.org/conf/2006/p9-halevy.pdf
  5. 5.Goh, C. H., Bressan, S., Madnick, S., & Siegel, M. (1999). Context interchange: new features and formalisms for the intelligent integration of information. ACM Transactions on Information Systems, 17(3), 270-293. https://doi.org/10.1145/314516.314520